You are on page 1of 72

School of Management and Economics

Demand Forecasting
- A study at Alfa Laval in Lund

Bachelor Thesis, FE 3583, 15hp Spring 2008 Authors: Hana Klimsova, 780329 Stacey Lobban, 740819 Tutor: Peter Berling Examinator: Petra Andersson

SUMMARY
Växjö University, School of Management and Economics, FE3583, Spring 2008 Authors: Hana Klimsova and Stacey Lobban Tutor: Peter Berling Title: Demand Forecasting- A study at Alfa Laval in Lund Background: Accurate forecasting is a real problem at many companies and that includes Alfa Laval in Lund. Alfa Laval experiences problems forecasting for future raw material demand. Management is aware that the forecasting methods used today can be improved or replaced by others. A change could lead to better forecasting accuracy and lower errors which means less inventory, shorter cycle times and better customer service at lower costs. Purpose: The purpose of this study is to analyze Alfa Laval’s current forecasting models for demand of raw material used for pressed plates, and then determine if other models are better suited for taking into consideration trends and seasonal variation. Delimitations: Due to the large number of articles that go into the production of plate heat exchangers, this study will only focus on forecasting of selected raw material (coils or sheets) demand for pressed plates. Method: This is an action research study using an interpretive scientific perspective, a deductive approach, and a quantitative research strategy. The scientific credibility of the study will hinge on providing a workable solution for the studied company’s problem. Conclusion: The company’s current error evaluation method differs greatly from theory and is probably wrong. Forecasting of raw material demand can be improved by using better input data. Of the 12 items tested, triple exponential smoothing provided the lowest forecasting error in 8 of the cases. Future Work: In our demand forecasting analyses, we adjusted each smoothing constant to come up with the best error accuracy for individual items. For future studies, one might use the same smoothing constants across all products to see if these can be standardized so that forecasting work can be done even more efficiently and effectively. Keywords: Alfa Laval, demand forecasting, time series analysis, moving average, exponential smoothing, autocorrelation, forecasting accuracy, plate heat exchangers

2

ACKNOWLEDGEMENTS

This Batchelor Thesis has been written during Spring 2008 and is a result of hard work and dedication. It would not have been possible without support and constructive feedback from a number of people. First of all, we would like to take this opportunity to give thanks and show gratitude to our tutor, Peter Berling, for his guidance and substantial feedback throughout the entire process. Furthermore, we are grateful for the additional guidance of our examinator, Petra Andersson, for her suggestions on how to structure our thesis. We would also like to express our gratitude to Ivan Kruzela, associate professor at Malmö University, for his conceptual and moral support. In addition, our appreciation goes to our opponents, who have given us constructive criticism and support in finalizing our thesis. We also appreciate the feedback of our fellow seminar members. Finally, our gratitude goes to all the people who took time to help us perform our research, including all interviewees at Alfa Laval OM-CP unit: Tobias Augustsson, Martin Jönsson, and Lars Hedberg. They provided essential value in the development of our thesis by sharing their knowledge and always made themselves available to answer countless questions and provide feedback throughout various stages of the process. Växjö, May 2008

Hana Klimsova

Stacey Lobban

3

......................................9 Disposition....................... 16 2...........................................2.................................. 14 2.....................................................2 Alfa Laval .......5 Purpose..............................................................................................1 Our data collection method................ 18 2.....................................company presentation................ 21 2....... 7 1...........................................................5 Data Collection.....8 Definition of Concepts...................................................................................................1 Our scientific approach .........................3 Problem discussion ...........TABLE OF CONTENTS SUMMARY..............................................1 Background........................ 17 2...................6 Delimitations ................................................... 23 4 ............................................................................ 16 2..........................................................................................3.........6.....................3 Research strategy – Quantitative vs.................................6 Scientific Credibility ......................................................................... Introduction .................Deductive ...7 Time plan ............................................................................. 2 ACKNOWLEDGEMENTS....... 7 1.1 Our perspective ....... 12 1......2 Scientific approach – Deductive vs..................................4 Problem questions ................................... Interpretivism ..................................................................4... 4 1............... 9 1.............................Theoretical and Empirical.................. 22 2.............................. Methodology ........5........ 20 2....................................................................................... 16 2.... 3 TABLE OF CONTENTS....................................................... 13 1... 11 1........................................................................................................... 17 2.......................... 13 1................ 12 1......... 12 1........... 18 2...............................................................................................1 Our motivation of research method ..........................................................................................Positivism vs............................................................................................................................................Action Research................................ 17 2.............................Quantitative ........................ Qualitative ............1 Our choice of scientific credibility ............1 Scientific perspective ................................................................... Inductive ...........1 Our research strategy ........................... 21 2...................................................Interpretivism .......................................................................4 Research method .............................................................................................1............................................

............. 42 4..............1......... 26 3................................................................................Subjective and Objective................................................2.......3 What the forecast sent to suppliers is based on .......................................................................................................................2 Exponential smoothing.... 49 5.................................................3 Ordering Raw Material at Alfa Laval...........2 Time series analysis..............................................................................................2 How forecasting of raw material demand is done .............................................................................................4........ Analysis .......... 47 5...........................1...........................................................1 Demand forecasting ...............2 Objective forecasting approach ......................3........... 40 4. 31 3.......................4 Forecasting accuracy calculations ..............................................................................1 Foundation of analysis ............................................................... 36 3.................................................................................... 50 5...................1 Subjective forecasting approach ................................................................................................................................ 42 4....3 Summary of subjective and objective approaches..................3 Autocorrelation ...........................1 Reintroduction of Alfa Laval ...................................................................................................4 Current forecasting models ...................................................................................................... 32 3......................................... 41 4..........4....... 24 3....... 43 4.................... 41 4................ 27 3..........................3............................................................4............................ 30 3...........2 Interviewees ..........................................1 How an order is placed ............3 The analytical process..........................................................5 Evaluating forecasting accuracy ...............................4 Models for Time Series Forecasting ............3 Definition of Time Series Behaviours . 28 3..................... 43 4....................... 25 3........................................................................................................................7 Summary of Methodology ..................................................3.......... 40 4..3.................................3............................................1 Moving average.................... Empirical........................ 28 3......1............................................ 49 5................ 41 4.....................................................................6 Results of poor forecasting at Alfa Laval...................................................................2 Raw material analyzed.......................... Theory .................................................................6 Gathering of data .................. 36 3.............................................................................................3............. 42 4...............1 Summary of the forecasting methods used by Alfa Laval.......5 Items to be analyzed ........................................................................................................... 46 4.............. 46 4........ 25 3.......................................................................................... 39 4....................................................................................................4............... 51 5 ........6 Theoretical framework ........................5 Six Sigma ......................... 25 3................. 44 4...................

.............................................................. 65 6... 63 6........................................................................................Historical data (one item) APPENDIX 3 and 4 .........................................2 Reflections .................................................... 64 6............................ 63 6.................................1 Results ................................................Interview questions APPENDIX 2 ............................................5. Conclusion......................................................................................................4 Future work .................................... 60 6. 66 APPENDIX 1 ............3 Client Validity.........................................................................................................Raw material analysis (one item) APPENDIX 5 ...................Analysis of evaluating method 6 ........................... 65 REFERENCES ............................4 Summary and discussion of analysis ...........................................................................

this chapter leads to a discussion of the problem. 1. K. Thereafter is a presentation of the company. capacity requirements…” to name a few. not everything can be accurately forecasted. It opens with a background of forecasting. (2007:85) 7 . Uncertainty and environmental turbulence are also issues to be considered. delimitations of the research. Subsequently. giving the reader insight to the important role of forecasting for demand. S. and key definitions. requirements and availability of raw materials. followed by problem questions. as there is a correlation between environmental turbulences and judgmental adjustments of quantitative forecasting techniques. sales of existing and new products. We make decisions about what to wear the next day or what to pack for a trip based on weather forecasts. The type of industry impacts the forecasting methods as well. Trends. product and market) compared to smaller ones. most often they are not.1 Larger firms employ forecasting for different purposes and for different levels (in terms of time. and process of this thesis. “All business planning is based on forecasts. cycles. (2005:53) Zotteri. “Although some portions of the demand process may be unpredictable.2 A company’s operations and production management’s main focus is to forecast product demand. Sometimes forecasts are accurate. Similarly forecasting plays an important role in business operations. Of course. motivation. other portions may be predictable.1 Background Forecasting Forecasting plays an integral role in our daily lives.1. Manufacturing firms use more complex techniques than service firms. Introduction T he introduction chapter is intended to provide an overview of the subject. and seasonal variation 1 2 Nahmias. We decide when to make investments based on forecasts or listen to traffic reports to determine which route to take to work in the morning. Even though it is very likely that demand will be random in most circumstances. forecasting methods are still of value.

S. Accurate forecasting is a real problem at many companies.6 To predict future demand accurately is crucial for effective purchasing planning since it leads to customer satisfaction and decreasing costs. A. (2005:51) 8 . extensive coordination is required among ordering. M. The managers here are aware of the fact that the forecasting methods they use today can be improved or replaced by others. Over-purchasing not only negatively affects inventory and storage space. & Touminen. Kaoien. Along with improving forecasting accuracy. P. R. (2007:1605) 6 Martin. (2000:4) 5 Kengpol.5 Good demand forecasting helps to find the right balance.” This statement is based on the fact that they too often experience unnecessarily high errors in their forecasts. 3 Such forecasting outcomes are important information for the use of purchasing raw material.7 Many times business managers make the comment: “If only we could do a better job of forecasting demand.. but also spending on unnecessary procurement activities.8 Improving forecast accuracy is the goal. however their forecasts are often adjusted by using personal judgment and market knowledge to improve accuracy.may be present” and will prove advantageous to making predictions of outcomes. but one must also take into consideration that there are limits.. Many companies invest in expensive forecasting software that use statistical methods.9 Alfa Laval also experiences the same problem of forecasting for future demand. under-purchasing can cause a shortage in raw material.4 The importance of demand forecasting manifests itself in the purchasing function and can have a strong influence a company’s competitive advantage. S. A change could lead to better forecasting accuracy and lower 3 4 Nahmias. The better forecast always includes some measure of the forecast error. On the other hand. J. (2007:16) 9 Nahmias. & Goodwin.. L. (2008:231) 7 Fildes. As goods move from raw material processors through to manufacturers and further down the supply chain to the end customer. D. K. there should be an overall strategy of demand management. & Zeng.. which could cause interruptions in the manufacturing process and lead to unfinished or delayed products. (2007:42) 8 Lapide. (2005:53) Fenstermacher. demand forecasting and inventory decisions at every level. This would include making an increasingly larger proportion of product demand firm while at the same time decreasing the percentage of products that have to be forecasted. P.

2008-04-06 Alfa Laval Annual Report 2006 9 . The company offers both standardized plates and those made according to customer specifications. chemicals and petrochemicals. titanium. Alfa Laval’s equipment.stainless steel.errors which means less inventory. of which there are several worldwide locations. where they are assembled. pharmaceuticals. the plates are put together with frames to form the final product – plate 10 11 http://www. systems and services are used by customers to optimize the performance of their processes. Its success was based on his invention of the continuous separators. India.2 Alfa Laval .Component Unit Plates). heat transfer and fluid handling solutions help customers’ heat. This unit is responsible for ordering and receiving raw material (a variety of metals . These factors will play an important role in the company gaining competitive advantages and flexibility. shorter cycle times and better customer service at lower costs. Almost 50 per cent of total sales are made in Europe. cool. Heineken and Tetra Laval. OM . Bayer. BP. 30 per cent in Asia and 20 per cent in North and South America. separate and transport products such as food and beverages. sugar and ethanol. This facility is comprised of mainly two divisions: CU (Component Unit) and SU (Supply Unit). Alfa Laval sells its products to more than 20.company presentation Gustaf de Laval founded Alfa Laval in 1883. USA and France. starch.CP (Operation Manufacturing . 10 Today the company has about 10.com. Over the years.11 Alfa Laval in Lund Alfa Laval’s facility in Lund specializes in the production of plate heat exchangers in different sizes and types depending on the purpose of use.000 employees of whom many are located in Sweden. This study will focus on one business unit. heat transfer and fluid handling has been added to the expertise in separation.) in form of coils or sheets to be cut and pressed into plates of different technical parameters and characteristics. 1.000 customers including BASF.alfalaval. These plates are delivered to the SU (Supply Unit). Both divisions are divided into two business units. Denmark. For example. etc. After that they are welded or outfitted with gaskets to become finished plates.

These are quality controlled and prepared for delivery to an end customer. Suppliers Component Unit . 10 .1: Alfa Laval’s Value Added Chain.heat exchangers. this study’s focus will be on unit highlighted in red.CP (Component Unit Plates) OM .SE (Supply Equipment) OM .SP (Supply Process Technology) End customer Figure 1. In the below value added chain.SU OM .CU OM .CF (Component Unit Frames) Supply Unit .

One method weights the previous 12-month demand to forecast demand for one year ahead.1. 2008-04-04 11 . The final forecasts generated by the ERP system have to be subjectively adjusted to meet actual demand. The further back in the history (always 12 months). This is time consuming and the human factor involved in this task can become a source of mistakes.3 Problem discussion As mentioned earlier. The forecasting methods were developed internally and resemble variations of moving average methods. In addition. forecasting errors are simply too high. In other words. the lower the weight factor is. The other method distributes the weight evenly on all twelve past periods. Today the OM-CP business unit uses mainly two types of forecasting methods to forecast demand of raw material to be purchased for the different families of plate heat exchangers. this study will focus on business unit OM-CP (Operation Manufacturing-Component Unit Plates). There are other problems as well. It is assumed that the forecast for the next three months will be more accurate if it is based on demand development during the preceding three moths. the automatic calculation of both demand forecasts has to be manually adjusted in order to take into consideration the value of known orders of the next months. the current weight factor methods seem not to account for trends and seasonal variations in demand.12 Alfa Laval’s current forecasting situation results in many questions: Is the manner of capturing historical demand sufficient to forecast future demand? Are there better forecasting methods to take trends and seasonal variations into account? Which methods work well and which do not in which situation? Which of the forecasting methods provides better results than the current method? 12 Interview with Martin Jönsson. However. the demand forecast for one year ahead is very similar to the actual demand in the past twelve months. There is an established rule that the biggest weight is put on the three preceding months. using this method. Generally. The OM-CP unit is experiencing problems with how to handle trends and seasonal variations in forecasting outputs. The monthly forecasting accuracy evaluation reports show negative trend caused by using these methods. The weight factor attaches different importance to each month.

we hope to suggest more appropriate forecasting methods for Alfa Laval. this study will only focus on forecasting of raw material (coils or sheets) demand for pressed plates (those classified by Alfa Laval as A or B products using the ABC classification system). 1. 12 .6 Delimitations Due to the large number of articles that go into the production of plate heat exchangers. 1.4 Problem questions • • How does Alfa Laval currently forecast for demand? How can the demand forecasting for raw material purchases be improved and which method/methods results in lower forecasting errors while taking into consideration trends and seasonal variations? 1.5 Purpose The purpose of this study is to analyze Alfa Laval’s current forecasting models for demand of raw material used for pressed plates.By evaluating the present situation and answering these and other questions. and then determine if other models are better suited for taking into consideration trends and seasonal variation.

1.7 Time plan Chapter 1 2 3 4 5 6 Title Introduction Research Methodology Theory Empirical Findings Analysis Conclusions and Recommendations Data Collection PM Tutoring 0 1 2 3 March Week 11 12 April 13 14 15 16 17 18 May 19 20 21 22 1. etc) to be cut and pressed into plates Sheet: raw material in form of flat metal plates Plate heat exchanger: The basic plate heat exchanger consists of a series of thin. The plates are then compressed together in a rigid frame to create an arrangement of parallel flow channels. titanium. welded together (or any combination of these) or brazed together depending on application. corrugated plates that are gasketed. One fluid travels in the odd numbered channels. 13 . the other in the even.8 Definition of Concepts Coil: a roll of raw material (metal.

1. It also presents the current forecasting techniques used by Alfa Laval and introduces the raw material selection for this study. we propose other forecasting methods for raw material demand taking into consideration trends and seasonal variations. Furthermore it presents methods used for data collection and empirical findings. and interviews with people involved in the OM-CP business unit. Recommendations for future studies give feasible ideas for studies of demand forecasting in this or another closely related subject area. The chapter concludes by summarizing the chosen method and discussing the validity and reliability of the study. Chapter 4 The Empirical Findings chapter describes the current order process and demand forecasting for raw material. 14 . Chapter 5 The Analysis chapter tests various forecasting methods using the historical data provided in Appendix 2.2): Chapter 2 The Methodology chapter describes the scientific research approach chosen for this study. Chapter 6 The Conclusion and recommendation chapter presents a review of the study and proposes forecasting recommendations to Alfa Laval. From theory and empirical evidence. Chapter 3 The Theory chapter provides relevant theories applied to our study. our data observations. At the end of the chapter we set up an analysis model based on applicable theories.9 Disposition The thesis will be organized as follows (see Figure 1. The proposals are based on theory. We will discuss different theories for forecasting of future demand using time series analysis.

Recommendations Chapter 6 Figure 1.2: Disposition of Study 15 .Introduction Chapter 1 Methodology Chapter 2 Theoretical Study Chapter 3 Interviews Empirical Findings Chapter 4 Internal Data Analysis of Data Chapter 5 Conclusions.

to closely collaborate with the practitioners to fully understand the meanings of their input and the empirical data gathered.”16 13 14 Bryman. (1998:80) 16 .13 “Practical action research seeks to improve practice through the application of the personal wisdom of the participants”14.scu.html (2008-04-15) 16 Greenwood. 2.au/schools/gcm/ar/arr/arow/rmasters. (1982:357) 15 http://www.Interpretivism We chose the interpretivist perspective because it is consistent with our research question of solving a practical demand-forecasting problem at Alfa Laval.Positivism vs. D.edu. the result also cannot take place without input from the practitioners. Areas covered include scientific perspective. S. At the end is a summary of methodological choices. and research strategy. Thereafter we present scientific credibility of the thesis. The implemented changes are naturally connected to the individuals involved in the change process because the result provides a solution to their immediate problem. Although theory plays a strong role in providing a solution. Practical action research encourages professional development by emphasizing personal judgment on decisions to act for the good of the client. A (2007: 19) Grundy. “The action research process thus creates a langue shared between practitioners and researchers that identifies the meaning throughout the inquiry process. scientific approach. Interpretivism Interpretivism is an alternative to positivism in that it requires the researcher to take into consideration the subjective meaning of social action.1 Scientific perspective . the researchers.15 2.1 Our perspective .2. Action research gains new understanding of a practice and tends to have a lasting influence on the changes that result. This requires us. Methodology I n this chapter we describe the methodology choices for this study. Positivism dictates that knowledge comes from strict scientific methods. Next the action research method is described and the manner in which data was collected.1.

and the researcher has the “technical skills in research procedures and comparative knowledge of the subject under investigation”. (2007:427) 21 Greenwood. The qualitative approach allows the subjects to play a greater role in designing the research and influencing results of the process.18 A deductive process is usually linear meaning that one step follows the other in a logical sequence. We will propose more appropriate forecasting models to solve the current problem. In quantitative research.3 Research strategy – Quantitative vs. the emphasis is on quantification of the gathering and analysis of data. Inductive research also involves an iterative process.2. For example. 2.20 The practitioners (those working at the organization) have the local knowledge and every day experience. E. A & Bell. Sometimes a researcher’s view of theory changes as a result of analyzing the collected data. This is a particular trait of action research. We will not create any new theories. Another differentiation between qualitative and quantitative research comes up from the relationship that qualitative researchers have with their research subjects. E. data may not fit the initial hypothesis or the relevancy of the data for a theory may become clear only after the data was collected. which requires the researcher to move back and forth between theory and data. A & Bell. (2007:581) Bryman. A & Bell.2. we do not know if we ought to test other forecasting theories. D.19 2. A & Bell. We have in mind forecasting theories that can be better suited to the current problem. E. However. (2007:14) 19 Bryman. data is collected to build theory. E. Inductive By using an inductive research approach. there are some situations where this does not happen.17 Deductive research seeks to apply theory to observations or findings. not test it. Qualitative Qualitative research places emphasis on words in the gathering and analysis of data rather than quantification.Deductive Our scientific approach is deductive in that we are analyzing forecasting models that are based on proven theory. (1998:77) 17 . but until all data has been assembled and we have tested the forecasting theory. (2007:13) 20 Bryman.1 Our scientific approach .21 Although action research is 17 18 Bryman.2 Scientific approach – Deductive vs.

3. Other information we obtained from the company’s intranet or website in addition to mostly unstructured interviews with three employees in the OM-CP business unit.Action Research A research method concerns choosing a technique to be used for data collection. E. E. E.web.26 Action Research.net/~robrien/papers/arfinal. A. & Bell. A. et al. & Bell.. (1998: 75-76) 18 .4 Research method .Quantitative Our research is mostly quantitative in terms of data collection and the analysis to be conducted. The researcher brings knowledge of theories and the responders bring their own personal experiences. (2007:429) 25 Bryman. requiring that we receive training and learn how to locate and sort needed information from the company’s ERP system.25 As a research method. preface x) 28 Greenwood. & Bell. participant observation recordings. this study also employs qualitative information gathering insofar as we are participating in our observations by gathering some of the necessary data ourselves. The diagnosis evolves through the close interaction of both parties. We are collecting and analyzing historical data of demand and using known theoretical formulas to find better solutions to the action research problem. a collection of documents related to the situation. (2007:40) 26 Bryman. or a diary of objective impressions24. et al. action research requires communication and collaboration between insiders (responders) and outsiders (the researcher) to diagnose a research problem and develop a solution.1 Our research strategy . C.. written description of meetings and interviews (can be given to participants to validate information gathered). it can employ the gathering of both qualitative and quantitative data. A. However. 2. D.likened to qualitative research. (1985. 2. A.html#_Toc26184652 (2008-04-15) 24 Bryman. case studies23. unstructured and structured interviews. (2007:428) 27 Argyris.22 Data collection can include such things as analyzing documents. E. & Bell. (2007:429) http://www. is a method that emphasizes combining science and practice27 and can further be defined as having the following characteristics:28 22 23 Bryman. in social science.

• • • AR treats the diversity of experience and capacities within the local group as an opportunity for the enrichment of the research-action process The meanings constructed in the inquiry process lead to social action. J. (1998:19) 32 Greenwood. (1998:54) 19 . to name a couple: 29 • • How can I improve my personal practice? How can I improve my understanding of this? This is because action research is a type of practitioner research that works as a tool to help improve professional activities in various types of workplaces. (1996:7) 31 Greenwood. J... D. (1996:16) Mc Niff.(1998:75) There is a special kind of question that action researchers ask in order to identify the particular research design.30 A quality criteria used to judge how good the theory is will focus on its ability to provide a practical solution in reallife situation. AR is inquiry where participants and researchers co-generate knowledge through collaborative communicative processes in which all participants’ contributions are taken seriously. D. or these reflections on action lead to the construction of new meanings. Figure 2.32 29 30 Mc Niff. et al. et al.Action Research (AR) – Core Characteristics • • AR is context bound and addresses real-life problems.31 Greenwood goes further to argue that “it is far more likely than conventional forms of social research to produce reliable and useful interpretations of social phenomena because the research validity can be tested in action. D. The credibility-validity of AR knowledge is measured according to whether actions that arise from it solve problems (workability) and increase participants’ control over their own situation.1: Action Research .Core Characteristics Source: Greenwood.

(1981:146) 2.Action research can be considered as a process that contains five phases: diagnosing. (1981:146) 20 . evaluating. G. has diagnosed a practical problem they are experience and we have the task of solving it.G.33 Figure 2. Alfa Laval.2: Cyclical Process of Action Research Source: Susman.4. 33 Susman.. Our work is characterized by finding a better solution for the current situation.1 Our motivation of research method The motivation for using Action research as a research method is because our subject company. action planning. and specifying learning. action taking.

some information pertaining to the demand forecasting and historical demand data will not appear in this paper. Some benefits to secondary data is that it reduces collection time and cost. (2007: 328-333) 36 Bryman.5. Vaxjo University library is the primary source for these media. It is raw data and classified in a way that meets the needs of the company. similar to what was mentioned under qualitative analysis. secondary data applies to previous demand forecasts and historical demand data taken from the subject company’s ERP system. is data already collected by the studied organization for its own purposes. R. Due to confidentiality concerns.. Primary data is collected from research. articles. interviews. 34 35 Yin. and also on the Internet. Therefore. direct observations. and E-brary to locate articles. no control over data quality and absence of key variables. participant-observation. and application of various forecasting theories. we have used the ELIN.Theoretical and Empirical Doing research requires the collection of different data sorts. Secondary data.34 Data may fall into the category of primary or secondary data collection. Our data collection will focus on books or peer-to-peer articles concerning purchasing. The empirical evidence. Theoretical data collection consists of information found in previous papers. (2003:86) Bryman.36 2.. information from different places in the ERP system have to be gathered and sorted by us as much as possible before it can suit the objective of the study and be used in our analysis. surveys.2. Here. and leaves more time for data analysis. content analysis.35 Some limitations include lack of familiarity with the data. conducted by the researcher for his/her own purposes. can come from documents. A.1 Our data collection method Our data collection method is composed mostly of secondary data. Besides books. theoretical and empirical. etc. complexity of the data. archival records. etc. A. (2007: 334-336) 21 . books.5 Data Collection. is this case.

(1996:108) 22 . This is vitally important to action research because of the collaborative nature of the research. If the method and its finding can be generalized to all situations. (1998:80-81) 40 Mc Niff. Second. then the research is proven to have external validity.2.38 Greenwood describes credibility in action research as “the arguments and the processes necessary for having someone trust research results”.37 Researchers share people’s knowledge collaboratively.2. evaluation of solution resulting into the modification of practice.6 Scientific Credibility The aim of all research is to create new knowledge. et al. The pattern of each cycle is the identification of an issue. (1996:106) ibid (1996: 107) 39 Greenwood. The stories of all people involved in a particular research are accumulated and demonstrate a culture of collective knowledge. J. implementation of solution. the criteria for high quality research are not appropriate for action research. there is external credibility – knowledge capable of convincing someone outside the project that the results are believable. 40 37 38 Mc Niff. nor desirable. J. gathering of evidence. imagination of solution.39 As seen in Figure 2. The replication or generalisation is not possible. These reports represent formative evaluations to check that one is in line with answering the research question. First there is knowledge that has internal credibility to the group generating it. However. Action research is not build of replicability and generalisability but of knowledge constituted of case studies. This results in the construction of collective knowledge. D. The aim is to understand rather than to predict. action research operates in cycles.. Features of traditional research are replicability and generalizability. It is recommended to make intermediate progress reports throughout each cycle. Research of high quality is characterized by meeting replicability meaning that it enables other people to do the same thing with the same results. The social intent of research is to improve a particular situation. Action researchers do research with other people to understand and contribute to improvement of their social practices.

Some forms of validation:41 Self - validation: Researcher should be able to show to his own satisfaction that he has been done the things he set out to do. Up-liner validation: To show to managers and authorities that an intervention has been done to improve the practice and the way of working could be adopted. Client validation: The intervention should contribute to improvement of quality of “client’s life”. Academic validation: This validation is done by the academic community in terms of whether it agrees that the researcher has contributed to a new knowledge.

2.6.1 Our choice of scientific credibility
This paper will use the form of client validation. Our recommendation should improve the current forecasting situation at the OM-CP business unit. To guarantee internal credibility, we will present to Alfa Laval a description of their current forecasting situation. On a regular basis, we will provide updates about our partial findings and progress and solicit feedback on our interpretation. All empirical findings and analyses will be documented and made available to the client in order to increase the credibility of our research. Collaboration between the company and us is critical to ensure our full understanding of the client’s current situation and demands. We believe that all these steps will lead to final knowledge that convinces observers outside the project that the study’s results are believable and useful in contributing to new knowledge.

41

Mc Niff, J. (1996:108)

23

2.7 Summary of Methodology Methodology in General
Scientific Perspective Scientific Approach Research Strategy Research Method Theoretical Data Collection Empirical Data Collection Scientific Credibility
Figure 2.3: Summary of methodology

Methodology in this Thesis
Interpretative Deductive Quantitative Action Research Articles, books, ELIN, Internet Unstructured interviews, secondary data, brochures Client validity

24

3. Theory

T

he theory chapter aims to present different methods for forecasting demand. We discuss two types of demand forecasting: subjective and objective. Next we identify behaviors that impact time series analyses. It is followed by several well-known

methods to evaluate forecasting error. Subsequently, a theoretical framework is provided to summarize the connection between the theories.

3.1 Demand forecasting - Subjective and Objective
“When the result of an action is of consequence, but cannot be known in advance with precision, forecasting may reduce decision risk by supplying additional information about the possible outcome.”42 There are two main approaches to forecasting, subjective or objective, and sometimes a company may opt to use a combination of both approaches to predict demand. It should be noted here that some authors use the terms objective and subjective43 while others use the terms quantitative and qualitative44, and still others use the terms judgemental and statistical.45 For the sake of consistency, we use the terms subjective and objective in this study. A discussion of both approaches follows.

3.1.1 Subjective forecasting approach
Subjective forecasting is based on human judgement. There are various types of subjective techniques:46 • Sales force composites – The sales force is in direct contact with customers so they have an advantage of having a pulse on consumer preferences. Based on this information members of the sales force can provide a sales estimate of products they plan to sell in the forthcoming year.

42 43

Kalekar, P., (2004:1) Nahmias, S. (2005:51) 44 Martin, J. (2008:222) 45 Armstrong, J. (2001:22) 46 Nahmias, S. (2005:55-56)

25

or bringing people together who have knowledge of a particular customer segment.48 The goal is to identify repeatable and predictable patterns in preferably several years of past data.50 An example would be predicting inflation (dependent variable) 47 48 Martin. (2005:51) 26 . the second is to hold a group meeting for the managers to come to a consensus. J. S. polling an organization’s sales force.2 Objective forecasting approach Objective forecasting is derived from analyzing data from past history and generally falls under two categories of analysis – time series analysis and regression models (causal).47 3. An example of time series forecasting is to predict sales for the next quarter based on sales from this quarter. subjective technique can encompass simply asking for managers’ opinions. if properly conducted. J.no group meeting . • Delphi method – This is similar to the jury of executive opinion in that it is based on expert opinions. It differs in that the managers are individually surveyed . One is for the person assembling the information to interview the managers directly. (2005:51) 49 Martin. • Time series analysis “uses only the past history of the series to be forecasted”. Therefore a model can be constructed to predict one phenomenon based on the development of other phenomenon. These opinions are collected from individual sources in several ways. (2008:222) 50 Nahmias.so that no personality can influence another in the group.1. (2008:222) Nahmias.• Customer surveys – These surveys or questionnaires. • Jury of executive opinion – If there is no past history data (such as for a new product) then expert opinions can be used as a source of information to derive a forecast. In short. S.49 • Regression models typically include the past history of other series. can indicate future trends and oscillating preference patterns.

org/smetoolkit/en/content/en/416/Demand-Forecasting (2008-04-29) 27 .g. etc (independent variables). 3.g. existing products. For instance.e.org52 Objective Approach Used when situation is stable and historical data exist (e. agricultural performance.51 Description Applicability Subjective Approach Used when situation is vague and little data exist (e. subjective approach is good when used to estimate a new product or service demand.1. current technology) Involves mathematical techniques Time series models Causal models 51 52 Martin.1). new products and technologies) Considerations Techniques Involves intuition and experience Jury of executive opinion Sales force composite Delphi method Consumer market survey Figure 3. however one need several years of demand patterns and it may encounter accuracy concerns.1: Summary of qualitative and quantitative forecasting approaches Source: SME Toolkit.3 Summary of subjective and objective approaches In summary one can say that each approach has unique advantages (see Figure 3. The objective approach is good when used to estimate demand of existing products. (2008:224) http://www.smetoolkit. This analysis model is further discussed under section 3. i.2 Time series analysis. In this study we will focus on the objective forecasting approach using times series analysis. a recession. It also provides an opportunity for the company to develop relationships with clients through closer communication. J. In addition.based on various factors such fiscal policies.. it is difficult to use when considering unusual events that impact customer demand.

Seasonality – Pattern that repeats at fixed intervals. time series methods are often referred to as naïve methods. inventory management. (2005:57) 56 Crum.. They are defined as:58 • • Trend – Refers to a tendency to show a stable pattern of growth or decline. distribution. However. Because it requires no other information than the past observations of the phenomena being analyzed. and its seasonality. (2003:32) 57 Lapide. L. S.. Some of these components are trend.1. (1999:29) 58 Nahmias..55 When demand is stable and few changes are expected to adjust the timing and volume of demand.2 Time series analysis As mentioned under 3. the time series analysis may be the most accurate input into demand planning. seasonality.53 Time series analysis better suits this study due to predicting demand for coming periods based on demand for same units in past periods.e. (2005:58-59) 28 . cycles and randomness.3 Definition of Time Series Behaviours It is useful when conducting time series forecasting to understand the causes of demand variations. Time series methods of analysis “assume that a time series is a combination of a pattern and some random behaviour. 3. (2004:2) 55 Nahmias. 53 54 Armstrong. which all depend on accurate forecasts.3. C. S. its long-term increase or decrease.”54 The goal is to separate the pattern from the white noise by understanding the pattern’s trend. ice cream consumption exhibits a seasonal pattern with increased demand during the summer. J. sales and marketing. i.2 Objective forecasting approach. time series analysis and regression models are used to analyze past series of data. P.56 This type of forecasting is also a necessity for companies that produce hundreds or more of items. budgeting. (2001:209) Kalekar.57 These demand variations can be broken out into components (otherwise called patterns or behaviours). time series forecasting has the largest number of applications of any approach to forecasting encompassing production planning.

” For example: a long-term economic business cycle that may be present in addition to seasonal fluctuations. “One can generate patterns purely at random that often appear to have structure. • Randomness .A series where data shows no recognizable pattern.• Cycles – “Variations similar to seasonality except the length and magnitude of the cycle may vary.” 29 .

. Trend plot Time series decomposition Plots time series data versus time without creating mathematical model Breaks a time series into its level. Models a level (stationary) time series with a trend and seasonality using three smoothing parameters. no trend or seasonality. Models a level (stationary) time series. Double exponential smoothing.3. trend. and Triple exponential smoothing.4 Models for Time Series Forecasting Below is a summary of major time series models and their applications. and irregular components. i. Source: Martin. Simple exponential smoothing.J. however. It models both trend and seasonal patterns using constants calculated from the decomposition. It should be noted that much more involved models exist.e. the main ones are presented below.2: Major Time Series Models Of these models. seasonal. trend. A time series model created by taking the average of observations from the time series to smooth out seasonal or other data patterns. using one smoothing parameter. the following appear in the empiric and analysis of this study: Moving average (and weighted moving average) models. 30 . (2008:240) Moving average models Simple exponential smoothing Double exponential smoothing (Holt’s method) Triple exponential smoothing (Winter’s method) Autoregressive integrated moving average (ARIMA) models Figure 3. and seasonal components of a time series. Models a level (stationary) time series with a trend but no seasonality using two smoothing parameters. Statistically based time series models that model level.

S. S. (2005:64) 31 . A small N puts more weight on recent observations. whereas a bigger N puts greater weight on past observations. (2005:64) 59 60 Nahmias.4.61 The N parameter (representing the number of periods) in moving averages is similar to the smoothing parameter of α used in exponential smoothing (discussed below in 3.3.Simple exponential smoothing). t ≥ 1) N = number of all past observations di = value of all past demand observations Source: Nahmias.1 Moving average Moving average is the simplest forecasting method. S. J.59 It is good at matching the average level of the time series period by period and it will follow the up.60 A common disadvantage of this method is that one must recalculate the average of the last N observations each time a new demand observation becomes available.4.2 . The moving average of demand N is calculated as the average of the most recent N data observations. Dt = observed values of demand during periods of time t (Dt.and downward movement of a time series. (2008:245) 61 Nahmias. (2005:63) Martin. Ft is the forecast made for period t in period t-1 and it is given by: t-1 Ft = (1/N)* Σ Di = (1/N)*(Dt-1 + Dt-2 + …+ Dt-N) i = t -N The formula represents a one-step-ahead forecast.

for lower 3. similarly.1 and 0.63 Forecast in period t : Ft = (α*(Dt-1+ …+ Dt-3) + β*(Dt-4+…+ Dt-6) + γ*(Dt-7 +. (2005:66) 65 Kalekar.4. S. This is good for short-range forecasting.Weighted moving average A weighted moving average adjusts the moving average method to reflect fluctuations more closely by assigning weights to the most recent data.0.2 generally represents a stable forecast and is recommended for production planning purposes.”64 It is a good technique for continually revising a forecast when taking into consideration more recent experiences. recent observations are given more weight in a forecasting than older observations. S. The higher the weight. In other words.com/308/forecasting-unweighted-and-weighted-moving-average-model. P.0 for a total of 1. 62 63 http://www. weights..+ Dt-12))/12 62 The weights lie between 0. (2005:69) 32 . usually a month into the future. Therefore. the more heavily recent demand will be weighted in a forecast. having a value of α between 0.0 and 1. then the higher importance we place on more recent data. the more heavily past demand will be weighted and forecasts are more stable.. 2008-04-29 http://shmula. “The current forecast is the weighted average of the last forecast and the current value of demand.66 A drawback to simple exponential smoothing is that it is not designed to forecast time series exhibiting trend or seasonal behaviours. The larger the α. Simple exponential smoothing Simple exponential smoothing constructs the forecast as a combination of previous demand weighted by a smoothing parameter α.65 The smaller the α. 2008-04-29 64 Nahmias.2 Exponential smoothing Exponential smoothing is another well-used forecasting method for stationary time series. (2004:3) 66 Nahmias.. meaning that the older data is usually less important.org/smetoolkit/en/content/en/416/Demand-Forecasting.smetoolkit.

3: Intercept and Slope Extracted from Lawrence. In other words. it is best used when data shows a pattern of trend.Formula: Ft = α*Dt-1 + (1-α)*Ft-1 That is: New forecast = α(Current observation of demand) + (1 . (2005:66) Double exponential smoothing (Holt’s method) Double exponential smoothing. The green line represents demand. (2008:23) A drawback using this analysis. level of the time series (expressed as the intercept) and trend (expressed as the slope). An example of the slope and intercept are shown at the tail end of the forecast. is how to first get the method started. Below is a figure demonstrating an intercept and slope. Sales Intercept Slope Figure 3. The best way to combat this is to establish a set of initial periods as a 33 .α)(Last forecast) Smoothing constant: 0< α ≤ 1 Source: Nahmias. Holt’s method. S. similar to simple exponential smoothing. is designed to track time series with linear trend. red represents forecast. It works like simple exponential smoothing except that two components. must be updated each period. S.

69 This method provides an advantage in that it is easy to update as new data is received. (2005:83) 34 . (2008:247) 70 Nahmias. 68 Triple exponential smoothing (Winter’s method) Winter’s method is built on the previous Holt method and is designed to handle time series base.html.70 The following model is assumed: Dt = (μ + Gt)*Ct + εt 67 68 Nahmias. trend. and seasonal behaviours by using three smoothing parameters: α to smooth base. When it is close to 0. t+η = St + ηGt St = α*Dt + (1-α)*(St-1 + Gt-1) Gt = β*(St-St-1) + (1-β)Gt-1 0< α ≤ 1 0< β <= 1 β≤α Source: Nahmias. S. S. and γ to smooth seasonality.starting point and use regression analysis to make estimates of the slope and intercept using the starting point data. J. Formula: η-step-ahead forecast made in period t Intercept: Slope: Smoothing constants: Ft.67 Also this method does not consider seasonal behaviours. S. more weight is given to recent observations (quick dampening).net/docs/net/sourceforge/openforecast/models/DoubleExponentialSmoothingM odel.0. β to smooth trend. 2008-05-23 69 Martin. (2005:77) http://openforecast.sourceforge. (2005:76) St = the value of the intercept at time t Gt = the value of the slope at time t Note: When the smoothing constant is close to 1.0. relatively less weight is given to recent observations (slow dampening).

one must obtain a building block of estimates for the series.73 Formula: Ft. The best assurance in this scenario is to choose smoothing parameters between 0. the current level of deseasonalized serie Gt = smoothes the trend forecast ct = smoothes the seasonal forecast 71 72 Nahmias. it is not the only way to start. however. it assumes t≤N St = smoothes the base forecast. (2005:83) ibid (2005:86) 73 ibid (2005:88) 35 .There μ is the base signal or intercept at time t = 0 excluding seasonality. Ct is the multiplicative seasonal component i period t.1 and 0. S. Gt is the trend or slope component. In addition. the slope and the seasonal factors. It is suggested that one must have at least two seasons of data to initiate the model. there is no assurance that the best values of the smoothing parameters based on past data will be the best values to use in future forecasts.72 Seasonal factors could be determined by a moving average method. and the error term is expressed as εt. (2005:84) Ft = the forecast made in period t for any future period t + η. S.2. or by using the slope and intercept values as proposed in Holt’s double exponential smoothing model. t+η = (St + ηGt)*Ct+η-N St = α*(Dt/Ct-N)+(1-α)*(St-1 + Gt-1) Gt = β*(St-St-1) + (1-β)Gt-1 ct = γ*(Dt/St) + (1-γ)*Ct-N Smoothing constant: 0< α < 1 0< β < 1 0<γ<1 Source: Nahmias.71 To initiate the model. A drawback to using this model is that the calculations become tedious since one has to test many combinations of the three parameters.

it is recommend to first test for autocorrelation. The formula for obtaining the forecast error for onestep-ahead forecast is following: et = Ft . (1983:55) 36 .arizona. (2005:91) 76 Hoff. S. seasonality. Dt-n) = COV(Dt. et is the difference between the forecast value for that period and the actual demand for that period. then the series exhibit a positive autocorrelation. -1 indicates strong opposite (negative) autocorrelation.5 Evaluating forecasting accuracy One should take precautions to conduct unbiased forecasts and one way to test this is through measurements of forecasting error. 2008-05-23 75 Nahmias. The value +1 indicates strong positive autocorrelation.ltrr.” A positive autocorrelation denotes the presence of seasonality and therefore triple exponential smoothing method is recommended.pdf . Autocorrelation is the correlation of a time series with its own past and future values. Forecasting error in period t.3 Autocorrelation Before using Winter’s method. “For example. A negative autocorrelation would denote the lack of seasonality. if low values in historical data tend to be followed by low values two periods later.74 In other words. Dt-n)/(STD(Dt)*STD(Dt-n)) COV = covariance of Dt and Dt-n STD = standard deviation of Dt and Dt-n Source: Hoff.76 Rn =COR(Dt. therefore triple exponential smoothing would not be recommended. J.4. At least 72 periods of observation are recommended in order to get a reasonable estimate using the autocorrelation function. (1983:280) 3.3. Under the right circumstances it may outperform simpler forecasting methods. 0 means no autocorrelation.183. 75 An autocorrelation is calculated as some value between -1 and +1.104/search?q=cache:aYWnjh43K6oJ:www.233.Dt 74 http://64.edu/~dmeko/notes_3.C. J. the autocorrelation function helps test for the presence of patterns.

(1983:263) 37 . MAD (mean absolute deviation) measures average absolute deviation of forecast from actuals. (1983:262) 79 ibid (1983:272) 80 Andersson.C. One of them is to graph the values of forecast error ei over time.78 n MAD = (1/n) ∑ |ei| i=1 2. G. S. MSE (mean squared error) measures variance of forecast error.There are three often-used measurements of forecasting accuracy – MAD. It is the average of the sum of the squared errors divided by n periods.81 n MAPE = [(1/n) ∑ |ei / Di|] x 100 i=1 As mentioned earlier.C. MSE. MAD is usually preferred because it does not require squaring. J. It is the average of all forecast errors divided by n periods. It is the average of all the ratios 100ei/Di. To that end. Another method is to sum all the forecast errors. there are different ways of tracking a forecast method. (1994:206) 81 Hoff.80 n MSE = (1/n) ∑ ei2 i=1 3. one should have an unbiased forecast. Although MAD and MSE are most commonly used. MAPE (mean absolute percentage error) measures absolute error as a percentage of the forecast. If the value of this moves too far away from zero (either above or below). Forecast errors should fluctuate randomly above and below zero if the method is unbiased. (2005:60) Hoff. and MAPE. where Di is the value of demand and ei is the forecast error at period i divided by n periods. it indicates that the forecasting method is 77 78 Nahmias. MAPE is a measurement “not dependent on the magnitude of the values of demand”. All errors are treated as a positive value. Σei. 77 1. J.79 This method puts more weight to large errors.

(2005:61) Andersson. G. S.biased. (1994:206) 38 . It is useful to use for comparing between different time series because the error is divided by corresponding demand value.83 82 83 Nahmias.82 MAPE puts less weight to large forecast errors than MSE.

3: Theoretical framework 39 .6 Theoretical framework To facilitate a clear understanding for our readers. we present the following model to visualize the connection between different parts of the forecasting theoretical framework. Demand forecasting Subjective (Human judgment) Objective (Mathematical) Time series methods Characterized by patterns: -Trend -Seasonal -Cycles -Randomness Regression (Causal) Characterized by dependent variables and independent variables Time Series Models of Analysis -Simple Moving Average -Weighted Moving Average -Simple Exponential Smoothing -Double Exponential Smoothing (Holt’s method) -Triple Exponential Smoothing (Winter’s method) Evaluating forecast accuracy -MAD -MAPE -MSE Figure 3. The highlighted areas directly affect our action research study.3.

The SU then produces the plate heat exchangers per client specifications and is responsible for quality control and subsequent delivery to the end customers. brewery. Several components such as stainless plates (i.) in form of coils or sheets to be later transformed into plates used for heat exchanger products. 40 . and heat transfer. The customers represent companies in various sectors such as hospitals. filtration. The OM–CP business unit is responsible for ordering and receiving these materials to meet customer demand.stainless steel. and forecast and error evaluation models currently used at Alfa Laval. T his chapter contains the empirical data collected throughout the action research study at Alfa Laval in Lund. Then we introduce the items to be analyzed and discuss how data was gathered. Purchasing and storage of raw material.1 Reintroduction of Alfa Laval Alfa Laval employs 10. the SU (Supply Unit).000 people worldwide. The facility in Lund.100 people and is responsible for the production of plate heat exchangers. fluid handling. titanium etc). marine etc. Empirical a short reintroduction of the company. etc. titanium. OM-CP’s manufactures the plate heat exchangers per orders from the internal customer. and production take mostly place at the facilities in Lund. oil and gas industry. demand forecasting. food industry. located at the same location in Lund. The company provides various products and solutions for centrifugal separation.4. followed by information about the current ordering process for raw materials. which fall under the category of heat transfer. The facility at this location is responsible for the production and storage of plate heat exchangers. We begin with 4. This study will focus only on the raw material demand of stainless steel plates (a variety of metals .e. Sweden employs 1. The questions can be seen in Appendix 1. frames and gaskets go into the production of these plate heat exchangers.: steel.

In other words. then the purchaser must place an order to one of the stainless steel suppliers (currently five or six available). a purchaser reviews the order in Jeeves. At OM-CP two purchasers are assigned to stainless steel products and they also handle product forecasting. and Lars Hedberg of the OM-CP unit throughout this study.4.3 Ordering Raw Material at Alfa Laval Alfa Laval has a plethora of product offerings. Below is a diagram overview of the order process. if the required raw material component is below the necessary stock level.2 Interviewees We received assistance from Tobias Augustsson. 4. the ERP system. In Jeeves it is recorded as quantity in kilograms and also converted into quantity of pieces for internal communication purposes. Martin Jönsson. An order handler keys the order into Jeeves.1: Current order and forecasting process 41 . This means that for all stocked items. these items of higher value will give more frequent procurement and therefore need more attention in inventory management. Purchasers are assigned responsibility for different plate types. Figure 4. The order quantity is communicated to suppliers in terms of kilograms. 4.1 How an order is placed The OM-CP unit receives internal orders from SU per customer specifications. A items represent 80% of the volume value. so for the purpose of this case we will only focus on the purchase of stainless steel coil that fall into the A or B category (using A-B-C classification). Upon review. Next.3. and B items represents 15% of the volume value.

Therefore.2 How forecasting of raw material demand is done Forecasting for all Alfa Laval suppliers is conducted once a month. if SU places an order for a plate product but this product is not in stock. The SU places orders to stock. However. The value is then divided by fifty (50 weeks/year) in order to get an average forecast for one week. the purchasers do not record their subjective forecast adjustments in the system.4. However. 4. 4. it is not possible for the database to capture a real history of how much the demand was adjusted up or down. The forecasting is based on data found in Jeeves and recorded as items “delivered from stock”. three-months and 12-months ahead. the one week forecast is multiplied by the number of weeks of the two nearest months and in the same manner also by the number of weeks in the third month.3.3. These manipulated (subjective) forecasts are sent to a supplier containing raw material demand one-month. Purchasers have to manually adjust (subjective forecast) this statistical forecasting every month to take into consideration known information such as current orders. the accuracy of measurement is based only on data registered in Jeeves. The adjustments are only recorded in excel spreadsheets. Updating of forecasts for suppliers is done on a monthly basis. “delivered from stock” is not necessarily a true representation of the demand since situations arise where orders gets placed. For example. Alfa Laval considers demand based on what is delivered from stock. However. In order to get forecast for the three following months (two and then one). however the measure of demand is calculated based on what has actually been delivered from stock.3.4 Forecasting accuracy calculations Forecasting accuracy (testing the level of forecasting error) is also updated once a month. 42 . which does not take into consideration manual adjustments of forecasts. The accuracy result measurement is currently about 50% correct so there is a lot of room for improvement. not in Jeeves. This value will later be used to calculate two and three month forecasts. but not fulfilled due to lack of stock.3 What the forecast sent to suppliers is based on A one year forecast is automatically generated in Jeeves using different forecasting methods. then demand is recorded as zero (0) because for the purpose of forecasting.

Under forecasting can lead to shortages in stock. the values falling within the limit are classified as “0” and those exceeding the limit as “1”. The margin is negotiated with each supplier. current forecasting does not show any improvement what so ever. which halts production because of lack of material. then most likely Alfa Laval has to follow through with the purchase even if that quantity is not required at present.usage one year ahead divided by four 4.5 Six Sigma Forecasting error is measured in the DPMO-measurement system (defects per million units) within the Six Sigma program. 43 . However.3. See Appendix 5 for further explanation and analysis of Alfa Laval’s accuracy evaluation method. Next. This value is then compared to real demand (delivery from stock) during the past 3 months. there is a certain margin built into the forecasting in terms of + and . However. This quality control also contains an assessment of how large a forecasting error has been made (the forecasting errors exceeding the limit of ± 30 % are registered in Six Sigma). It should have an accuracy of ± 30 %.quantity. Unfortunately.6 Results of poor forecasting at Alfa Laval If a forecast is too high and the supplier produces the volume.sum of the past three months F3m = Forecast . 4.3. this is a very rare occurrence. There is a limit for accepted forecast error. the error is expressed as a percentage of the forecast. Incorrect forecastings are at 50%. There is no universal margin across all suppliers.Calculation: Each month is the forecast value for the next 12 months divided by 4 in order to get a 3 month forecast. In Jeeves. Alfa Laval’s formula for estimating the forecast error as a percentage (%): et =((D3m – F3m)/F3m)*100 et = forecast error in period t D3m = Demand (Delivery from Stock) .

75 44 .4 Current forecasting models Alfa Laval has a selection of four different forecasting models.9 δ = 0.. We have found no evidence of this method being used. The four forecasting models are: 1.+ Dt-12))/12 Dt = Demand (Delivered from Stock) in period t Weighted factors: α=2 β=1 γ = 0.+ Dt-12))/12 Weighted factors: α = 1. the products manufactured at the facility located at Öresundsväg were forecasted using a method named “Ö-väg method”. products manufactured at a facility in Gunnesbo used “Gunnesbo method”. This manner of using a model based on where the production takes place is no longer in use.. Forecast in period t: Ft = (α*(Dt-1+ …+ Dt-3) + β*(Dt-4+…+ Dt-6) + γ*(Dt-7 +.... only that the methods seems to be appropriate for current demand forecasting needs. Forecast in period t: Ft = (α*(Dt-1+ …+ Dt-3) + β*(Dt-4+…+ Dt-6) + γ*(Dt-7 +. Öresundsväg (ö-väg) – resembles weighted moving average in which the most weight is placed on the last three months. meaning that current employees don’t really know when or how they came into being. However. In other words.4.+ Dt-9) + + δ*(Dt-10 +. Two of them are named after facilities located in Lund.10 γ = 0.5 2.25 β = 1. In the same manner. these forecast models have been grand-fathered into the system.. Gunnesbo (G-BO) – weighted moving average – in the same manner as “Öresundsväg method” the most weight is placed on the last three months.

1 However. Exponential smoothing – the forecast in period t is the weighted average of the forecast in period t-1 and the current value of demand in period t-1. Ft = α*(Dt-1+ (1. 45 . Note: Theory dictates that the smoothing constant values should be >0 and <1 and that the average of weighted factors should be 1. Forecast in period t: Ft = (α*(Dt-1+ … + Dt-12))/12 Weighted factor: α=1 4. RAK – straight moving average – the weight is distributed evenly on all the twelve past periods.α)*Ft-1 Smoothing constant: α = 0.3. this method. though available in Jeeves is not used.

As mentioned earlier.90 0.75 0.5 0. they are replaced by numbers of convenience.5 Items to be analyzed Below is a summary of selected raw materials and the corresponding forecasting method used at Alfa Laval. (see Figure 4.10 1. 46 . -12 represents 12 months ago.0 2.25 1.0 2.4. Exponential smoothing is not listed as we saw no evidence of its use.75 0.90 0.5 0.0 2.0 1.75 0.0 1. Each column to the right of Period represents a weight assigned to that period.5 0.5 1.4.25 RAK Weighted factors 1 1 1 1 1 1 1 1 1 1 1 1 Öresundsväg (ö-väg) Weighted factors 0. Alfa Laval suggested these items for the purpose of conducting the analysis.5 0.90 1. To respect Alfa Laval´s confidentiality agreement. Jeeves.2: Summary of the forecasting methods used by Alfa Laval 4. Data was obtained from the ERP system.10 1.0 Figure 4.1 Summary of the forecasting methods used by Alfa Laval Note: -01 represents the most recent past month.2) Period -12 -11 -10 -09 -08 -07 -06 -05 -04 -03 -02 -01 Gunnesbo (G-BO) Weighted factors 0. Instead.25 1. original item names and numbers will not be presented in this study. these items were selected based on high order quantity and value.10 1.5 0.

3: Current forecasting methods for selected raw material items 4. the others use “Rak method”. 47 . data we extracted from Jeeves and transferred to Excel sheets) and primary data – information gathered from unstructured interviews.6 Gathering of data There are 34 raw material items to be analyzed in this study.Raw Material Item Number 3 105 208 310 413 487 590 693 796 899 973 1014 1117 1220 1266 1369 1473 Current Forecasting Method Rak Ö-väg Ö-väg Ö-väg Ö-väg Rak Ö-väg Rak Ö-väg Ö-väg Ö-väg Ö-väg Ö-väg Rak Rak Ö-väg Rak Raw Material Item Number 1577 1681 1784 1806 1827 1906 2009 2103 2207 2229 2332 2410 2513 2591 2713 2816 2846 Current Forecasting Method Ö-väg Rak Ö-väg Rak Rak Ö-väg Rak Rak Rak Rak Ö-väg Ö-väg Rak Ö-väg Rak Rak Rak Rak: Straight Moving Average Ö-väg: Weighted Moving Average Figure 4.e. We received training to use Jeeves to extract past demand history for the raw material items. which influence our approach on how to analyze the data. The data for this study is a compilation of secondary data (i. Approximately half are currently being forecasted using the “Öresundsväg” forecast.

it was of great importance to identify patterns (seasonality.The historical data we gathered and sorted are presented in Appendix 2 and have the following characteristics: • • • • Observations of only one phenomena – demand of raw material items based on past values of “Delivered from Stock” Historical data is chronologically sorted (ie: 2001 . trend…) in data.2008) The quantity of available past data varies (ie: varied numbers of observation) Several zero (0) values are presented (ie: recorded as “0” delivered from stock) After the historical data for demand of raw material was obtained and chronologically sorted. 48 . This will be discussed further in Chapter 5.

demand is measured as delivered from stock. time series analysis is a necessity for companies that produce hundreds or more of items. The data and the complete analysis of all items are included on the CD. 3 and 4 in this report will present the analysis of one raw material item only to illustrate our approach. If the correct amount demanded is not available. then “0” product is delivered from stock. it means that technically the actual demand is not captured. when in actuality there is a demand. In other words. It also prevented us from being able to evaluate some of the other selected coils as many periods showed a zero value. This caused some difficulty in the analysis and forced us to consider using different methods of analysis better suited for the perception of “0”demand. thereby reducing the number of historical periods available for evaluation making time series method of analysis inappropriate for some coils. This poses a challenge when evaluating demand this way because we can only conduct an evaluation based on what has been shipped from stock causing a blind spot that makes it difficult to perform a proper analysis of periods where no demand was recorded as shipped from stock. the Appendix 2. Analysis I n the analysis chapter we use various forecasting theories to analyze demand for raw material items presented in the empirical chapter. our evaluation of such coils was hampered in that the periods of “0” value sometimes spanned a significant number of periods. 5. Both criteria are suitable for the raw material analyzed in this study. As mentioned in the theory chapter. 49 . We will seek to offer options of theoretical model that can provide better forecasting accuracy in practical applications. In addition. Note: Due to the huge amount of data produced during our study. At Alfa Laval.1 Foundation of analysis Our analysis was based on time series methods because the objects of our study have historical observations of repeatable demand that span several years. but it was not captured in the system because nothing was delivered from stock.5.

in the following) to evaluate.Nine of them have all past demand values available and three contain “0” values in the original data.5.2 Raw material analyzed We had a selection of 34 coils (raw material items.1).1: Raw material item numbers and current forecasting methods 50 . and the “0” value problem. Based on the time required to fully analyze each coil. our study refocused on the analysis of 12 of the raw material items (see Figur 5. The remaining items contain too little historical data or too many zero demand values and therefore they are not appropriate for analyzing with time series forecasting methods. The following 12 raw material items were analyzed: Raw Material Item Number 3 105 208 310 487 590 693 796 1117 1266 2410 2513 Current Forecasting Method Rak Ö-väg Ö-väg Ö-väg Rak Ö-väg Rak Ö-väg Ö-väg Rak Ö-väg Rak Figure 5. the behaviour exhibited by some coils during our analysis.

(See the note in the beginning of this chapter. these periods were corrected to an average of the before and after period.3 The analytical process Following is the step-by-step process we took to analyze the raw material items. or as a multiple-period average.2: Historical demand data . Appendix 3 contains complete analysis of one raw material item.5.) Step 1 Demand data was gathered and sorted in chronological order from Jeeves database.2 and also Appendix 2 for complete list of historical demand data for one item). The reason for correcting “0”periods is because they would otherwise disrupt the formula for 51 . Figure 5. Demand is based on delivery from stock (see Figure 5. A snapshot of the physical analysis accompanies a written description.snapshot Step 2 Where demand is positive in most periods and there were some “0”periods.

a diagram was created of the demand data to show if a trend or season in demand exists. Another reason is that the time series methods require more or less constant demand.3: Historical demand with corrected values Step 3 After each raw material item was scrubbed and “0” periods corrected.4: Delivery from stock (raw material item nr. Figure 5. See enclosed CD for complete data for all raw material items. this is a good indication that Holt’s method ought to be used. If a trend exists. 3) with increasing linear trend 52 .4) Delivery from Stock (Item Nr. (see Figure 5.computing the forecasting accuracy (MAPE).3) 60000 50000 40000 Pcs 30000 Delivery from stock Linjär (Delivery from stock) 20000 10000 0 1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97 Months Figure 5. the new demand data to be used for analysis appears under the column Data Correction. As seen below.

All the six forecasting methods are programmed in excel to work automatically. Simple exponential smoothing. initial values are required. • To start the Single and Triple exponential smoothing methods. Double exponential smoothing. and γ) have to be tested in order to get the demand forecast with the lowest error. a moving average. Recall that Rak and Ö-väg are methods currently being used at Alfa Laval. β. the parameter and number of periods (n) have to be adjusted. and Triple exponential smoothing (see enclosed CD). Ö-väg.5. If required. 53 . we take the demand average of two separate seasons (V1. A snapshot of one of the analyzing methods. Each of the twelve raw material items is analyzed by each of the six methods: Moving average.Step 4 We then programmed time series methods formulas and autocorrelation into an excel spreadsheet. some adjustments have to be made in order for the methods to perform correctly for each specific item. V2). and for triple exponential smoothing. of one raw material item is in Figure 5. To start the value. the different parameters such as smoothing constants (α. for single exponential smoothing. Such adjustments are: • The demand data must be pasted to the column “Delivered from Stock”. This applies to each method. Rak. However. we take the average of the ten first demand observations. • After the initial data are selected.

With autocorrelation we test if a relationship between the data exists. By removing the trend from data we can see if there is some other pattern left such as seasonality or just pure randomness (See Appendix 3).Figure 5. the correlation should not be 0. 54 . Then Winter’s method can be recommended. we use an autocorrelation test of the data looking for evidence of seasonality.5: Snapshot of moving average analysis Before triple exponential smoothing (Winter’s method) is applied. If seasonality exists.

and MSE to determine the lowest forecast error of each method for each item.step ahead moving averages in order to determine which of them gives us the lowest error. we tested different n . double exponential smoothing. MSE and MAPE both on one-month and three-month bases (see Figure 5. MAPE.6). Since the moving average analysis doesn´t require a smoothing constant. MSE and MAPE in double exponential smoothing analysis 55 . we then had to adjust alpha. and gamma (smoothing constants) and use MAD.Step 5 Evaluating forecast error As mention above. beta. Ö-väg and Rak methods we calculate the value of MAD. depending on the method applied. Figure 5. To evaluate the single.6: Snapshot of MAD.

5.5 680 0. we divide the one-year ahead forecast by 12 in order to get forecast in period t. Theory based error evaluation (3 months): The last alternative is the same as the previous but calculated on a three month basis.MAPE (/D) and MAPE (/F)). (See Appendix 3 .5 1 Figure 5. The figure is explained below.not by the demand value as the common practice dictates. 2. *When Alfa Laval evaluates the forecasting accuracy. Theory based error evaluation (1 month): In the second alternative. Next the forecast is compared to the real demand in period t. This value is then compared to delivery from stock values (real demand) during the past 3 months.7. 3.Evaluating Alfa Laval´s forecasting methods (Ö-väg and Rak) We used three alternatives to evaluate Alfa Laval´s forecasting methods: 1. The result of using this method is demonstrated in Figure 5. In Appendix 5 we provide a detailed mathematical analysis of the methods in order to demonstrate the incorrectness of this method. Next.7:Forecast accuracy. 01/ 2007 Extracted from Alfa Laval´s database 56 . This is the reason why we compute the forecast accuracy using another method (explained below) based on the theory as presented in the Theory Chapter 3.842865 3647. the error is expressed as a percentage of the forecast. Item_no Fct_usage_one_yr_ah_DIV_4 sum_three_month compare diff Late XXX 4327. the forecast error is divided by the value of the forecast . *Note: It seams this evaluation method is probably incorrect. Alfa Laval’s internal error evaluation method for forecast accuracy: Each month is the forecast value for the next 12 months divided by 4 in order to get a 3 months forecast.

usage one year ahead divided by four Using the values presented in Figure 5.5)*100 = .84%.7. But in fact. The forecast error should be divided by the demand not by the forecast. we obtain the following result: et = ((680 . the true error is actually much higher using the common practice of error evaluation as presented in the Theory Chapter 3.3. in order to estimate how high the error is compared to the real demand.5)/ 4327. This example goes to prove that Alfa Laval´s method for evaluating error results in very low forecasting errors compare to errors obtained using the common practice of evaluating errors (see also Appendix 4).84% It means the error compared to the forecast is .As previously mentioned in the Empirical Chapter 4. 57 .5. Alfa Laval evaluates the forecast accuracy (%) using following formula: et =((D3m – F3m)/F3m)*100 et = forecast error in period t D3m = Demand (Delivery from Stock) .sum of the past three months F3m = Forecast . as shown here: et = ((680 .4327.5)/ 680)*100 = .4.536% Both alternatives are presented in our analysis.4327.

Item Nr.8 and enclosed CD) We also graph the forecast error over time in order to determine if the method is unbiased. we created a graph showing both demand and forecast in order to visually asses how closely the forecast followed demand. (see Figure 5. the method is unbiased when the forecast error fluctuate randomly above and below zero. Forecast made by double exponential smoothing (2exp) 58 .Step 6 For each analysis. As we explained in the Theory Chapter 3. 3 (2exp) 60000 50000 40000 Delivery from Stock Pcs 30000 Forecast Linjär (Delivery from Stock) 20000 10000 0 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94 97 Months Figure 5.8: Graph showing forecast and demand in upward trend.5.

another of the evaluations of accuracy (MAD. MAPE. MSE) may be more appropriate to Alfa Laval. the orange highlighted value is our observation of the lowest forecasting error. we then created a summary chart to evaluate error accuracy of all methods applied.Raw Material Analysis (Analysis table).9 below. In Figure 5. depending on other factors in a real life situation.9: Analysis of raw material items 59 . Note: 3 months MAPE evaluation is provided for comparison since it is a measurement used at Alfa Laval.Step 7 After a raw material item was analyzed using each method. See Appendix 4 . Figure 5. However.

9 .γ =0.γ = 0.β = 0.2 .9 .1 .1 9202567 2182 90 37 4843015 1442 71 1522779 844 57 13756486 3012 41 3588770 1451 30 940763 655 36 5726137 1936 28 1615093 983 21 395257 434 21 α = 0.1 5per Mav 12per Mav 5per Mav α = 0.9 6504583 1726 39 19947452 3297 40 6031122 1853 29 α = 0.9 .1 14 per Mav 12 per Mav 5 per Mav α = 0.β = 0.β =0.β = 0.7 .β = 0.8 .9 α = 0.9 .β = 0.5.10: Summary of the raw material items and the best forecasting methods 60 .4 Summary and discussion of analysis Based on our analysis.1 .β = 0.γ = 0.2 .γ = 0.1 . Forecasting methods Item Nr. 3 Forecast Accuracy Evaluating MSEn MADn MAPEn % 105 MSE MAD MAPE % 208 MSE MAD MAPE % 310 MSE MAD MAPE % 487 MSE MAD MAPE % 590 MSEn MADn MAPEn % 693 MSE MAD MAPE % 796 MSE MAD MAPE % 1117 MSE MAD MAPE % MAPE % (3 months)/D 1266 MSE MAD MAPE % 2410 MSE MAD MAPE % 2513 MSEn MADn MAPEn % 5per Mav 5per Mav 5per Mav 330198 418 49 α = 0.β = 0.9 Parameters Moving Average 77172831 6554 32 Double exp smoothing Triple exp smoothing Rak method Ö-väg method 228272046 99077664 12503 7992 19 12 95752687 235368856 1.1 .9 . below is summary chart of lowest forecast error results for the 12 raw material items.γ = 0.γ = 0.9 .1 α = 0.7 α = 0.γ = 0.1 .β = 0.7 α = 0.1 .2 .γ = 0.12E+08 7517 11858 8121 33 39360646 4214 36 24968106 3237 39 34870514 4670 39 16 11 66815561 31394540 6266 4319 15 10 36908608 15783089 4382 2860 17 11 78422544 35771303 7278 4968 19 14385157 2275 16 13 6199812 1544 11 40283391 17921812 5378 3597 21 15 31332630 13594254 4556 2984 22 21329880 3515 28 14 8977227 2339 19 Figure 5.

4. one would expect that the best forecasting method should be determined with the help of following rules of thumb: 1) The moving average and single exponential smoothing forecasting methods should be used if the data exhibits no trend. Using this method guarantees the positive forecast values even if the forecast error is very high. 61 . which had to be adjusted manually by averaging observations. the forecasting methods providing the lowest errors are Alfa Laval´s methods (Rak and Ö-väg). Trying different smoothing constants in order to eliminate the negative values doesn’t help the situation. five periods moving average gives the lowest forecast error for items 3.According to the summary above. 90% (MAPE). This negatively impacts the performance of the forecasting methods. Items 1266 and 2410 exhibit very high forecast error compared to other items. 590 and 2513 despite the fact that all of them exhibit slightly increasing trend. The reason for this result is that the evaluating method of forecasting accuracy employed by Alfa Laval is not correct. reveals the fact that other time series analysis methods provide better results. Moving average is supposed to lag behind the trend. However. In the summary above. We decided to manipulate smoothing constants at double exponential smoothing instead. (See Appendix 5). However. 3) Triple exponential smoothing is an appropriate forecasting method if the data exhibits seasonal pattern with or without trend. Using triple exponential smoothing gives low error by 35% (MAPE). the relatively low error leads to negative forecast values. They have many zero values. Item number 1117 exhibits a lot of varying demand over time. Using the evaluating method. which are not desirable. The reason is the character of the historical data. (See Appendix 3 and 4) According to the theory presented in Theory Chapter 3. the result of analyzing the 12 raw material items using time series forecasting methods is in contrast to our expectations. corresponding to the common practice. 2) Double exponential smoothing (Holt’s method) accounts for a trend in the time series data.

Altogether we tested six forecasting methods. The only method that seems to have the ability to cope with our data is triple exponential smoothing. γ).2. This probably makes the method more flexible to provide better results than the other methods. which states this method is successful with data that has seasonal pattern. Our historical data. which is not our case. based on delivered from stock values. We carefully examined the available historical data spotted by graphing in order to answer the following question: “Why don’t these forecasting methods work according to theory?” According to theory. on 12 raw material items. inclusive of Alfa Laval’s methods. β. which can be manipulated in order to get the lowest forecast error. 62 . This result contradicts theory. the time series forecasting methods require more constant demand over time. The combination of high value of α and law value of β provides the best result. presented in Theory Chapter 3. in most of the cases don’t comply with this requirement. The method is based on three smoothing constants (α. The triple exponential smoothing method provides the lowest forecasting error in 8 out of 12 cases.

Then we provide 6. The disadvantage of using these models is that they are not the best tools to consider trends or seasonality.which resembles Straight Moving Average. Additionally. which are exhibited in some of the raw materials analyzed. In order to get the most reliable result of the forecasting method. T his chapter concludes our findings and answers our research questions. Finally are our reflections concerning the writing of this thesis. Alfa Laval divides the forecast error by the value of the forecast . we have found evidence that Alfa Laval currently uses two of four internally devised forecasting models. 1) How does Alfa Laval currently forecast for demand? For the selected raw materials evaluated in this study. and Ö-väg . since the formula is comparing the future data with the past data.6. client validity. The two models are internally referred to as Rak . and then some observations on conducting the data analysis.1 Results We begin this chapter with a reminder to the reader of the research questions posed at the beginning of this thesis and the response. This lack of adequate measurement contributes to the 50% accuracy Alfa Laval is currently experiencing. it was observed that Alfa Laval’s current error evaluation method differs greatly from theory and is probably incorrect. We address the validity of our research and give recommendations to Alfa Laval. 63 .which resembles Weighted Moving Average. Our recommendation is to compare the demand data at period t to the forecast data at the same time period t. Furthermore. the forecasting error should be related to the demand instead of to the forecast. and future work.not by the value of demand as theory and common practice dictate. Conclusion suggestions for continued research.

it does not mean there weren’t any orders placed in that particular month. all forecasting is based on the Delivery from stock values. Doing so leads to some periods (months) with zero values in the input data. 64 . Because Alfa Laval records historical data in Jeeves in a way that suits their own business and administrative needs and not to aid forecasting. To make the analysis as simple a process as possible we tried to express all formulas in a “user friendly way” before doing the analyses so that only historical data and other parameters needed to be pasted in and adjusted as necessary. it was time consuming to make data suitable for the purpose of this study (for example: current forecasting error per each raw material studied). This was a quality control effort to make sure that opportunities for human error were at a minimum.2) How can the demand forecasting for raw material purchases be improved and which method/methods results in lower forecasting errors while taking into consideration trends and seasonal variations? The demand forecasting for raw material purchases can be improved by using other input data than Delivery from Stock. we tested four standard forecasting methods as well as two of Alfa Laval’s own methods on 12 raw material items. collecting the information.2 Reflections A great amount of time was required to create Excel spreadsheet functions to be able to use the forecasting methods. triple exponential smoothing seemed to provide the greatest flexibility. However. However. 6. The “zero problem” will be solved at most of the cases using placed orders as a source of input data instead of Delivery from stock values. The triple exponential smoothing method provides the lowest forecasting error in 8 out of 12 cases. The zero values signify there is no raw material delivered from stock. As mentioned in the analysis. and then sorting. recalculating and analyzing it. If the analysis had been based on different input. this is based on the Delivered from Stock data. Currently. It required looking in different places of the ERP system. other methods could have been more appropriate. This will also lead to better knowledge about the real demand. but in this situation.

3 Client Validity The functions programmed in the Excel spreadsheets. For future studies. 6. It is a usable tool for Alfa Laval for future forecasting and allows them to test different models for suitability depending on the raw material demand they wish to forecast. The Excel spreadsheet functions can also be used to forecast other products in Alfa Laval or in other external organizations that wish to forecast and evaluate error accuracy for products that can use time series related forecasting methods for historical demand data. were designed to make it simple to use and all the user needs to do is to paste in necessary data.6. as mentioned earlier. we adjusted each smoothing constant to come up with the best error accuracy for a particular item.4 Future work In our analysis of the raw materials. it could be a good idea to use the same smoothing constants across all products to see if these can be standardized. 65 .

. Studentlitteratur. Hoff John C. “Principles of Forecasting. Oxford University Press. Eller College of Business and Public Administration. M. Kluwer Academic Publishers. J.. E. K. & Morten. Financial Management Caspian Publisher.. Yin. D.. Smith.. Crum. L. D. S. Putnam. “Introduction to Action Research”.. Deakin University Press. 1998. Geelong.. “Demand Management Best Practices: Process. “Know Your Supply Chain”. R. 2007 66 . McNiff Jean. Goodwin. Articles: Fawcett. & Fawcett. Kaoien.REFERENCES Literature: Andersson. 2005. McGraw-Hill Irwin. 1983 Nahmias.. J. “Three Modes of Action Research”.. 1985. 2003. A.. 2001. P. “Case Study Research: Design and Methods”. “Action Science Concepts. ”Regressions. 2007 Kalekar. & Touminen.och tidsserieanalys”. & Bell.. 2007. “You and Your Action Research Project”. Publishers. & Ågren A. Methods. 1996. S. 2003. C. Sage Publications. P.C.. Ross Publishing..“Business Research Methods”. Grundy. Wadsworth.. Armstrong. A. S. University of Arizona Fildes. “A Practical Guide to Box-Jenkins Forecasting”. “A Procurement Planning Improvement by Using Linear Programming and Forecasting Models“. G. 1995 Fenstermacher. D. “Production and Operations Analysis”. S. ”The firm as a value-added system: Integrating logistics. 2004 Kengpol.. A Handbook for Researchers and Practitioners”. Principles. R. E-books: Argyris. operations and purchasing”. R.. Jossey-Bass Inc. & Zeng. Sage Publications. and Collaboration”.. Greenwood.. 1982. and Skills for Research and Intervention”. PICMET. P. & Jorner U. “Forecasting”. “Time Series Forecasting Using Holt-Winters Exponential Smoothing”. December 6. Routledge Falmer Publisher. 1994 Bryman.

2008 Susman.Core Characteristics Figure 2.ltrr. “Forecasting practices: Empirical evidence and framework for research”. Zotteri. 1981.104/search?q=cache:aYWnjh43K6oJ:www. The Journal of Business Forecasting.2008-05-23 Demand Forecasting Time Series Models..scu. Graceway Publishing Comapny. Kalchschmidt. L. The Journal of Business Forecasting. International Journal of Production Economics. M.. power point presentation by Professor S. 2008-05-23 http://64. Winter 1998-99 Lapide L. “Operational Excellence: Using Lean Six Sigma to Translate Customer Value through Global Supply Chains”.. 27... February 2007 Martin.edu/~dmeko/notes_ 3. “Don´t just measure forecast errors”.Lapide.1: Alfa Laval’s Value Added Chain Figure 1. G. January 2007 Other: www.233.au/schools/gcm/ar/arr/arow/rmasters.com http://www.3: Summary of methodology Figure 3.arizona. Lawrence.sourceforge.net/docs/net/sourceforge/openforecast/models/DoubleExpone ntialSmoothingModel.2.edu.html.web. Management Science. University of Colorado.html#_Toc26184652 2008-04-15 http://www.183.pdf . “Planned Change: Prospects for the 1980s”.net/~robrien/papers/arfinal.1: Summary of qualitative and quantitative forecasting approaches 67 .1: Action Research .smetoolkit. College of Business Administration. J.org/smetoolkit/en/content/en/416/Demand-Forecasting2008-04-29 http://shmula.2: Cyclical Process of Action Research Figure 2.com/308/forecasting-unweighted-and-weighted-moving-average-model. 200804-29 http://openforecast..2: Disposition of Study Figure 2. G.alfalaval. Vol. No. USA Alfa Laval Annual Report 2006 Interviews with Martin Jönsson and Tobias Augustsson at Alfa Laval Figures: Figure 1.html 2008-04-15 http://www. “New Developments in Business Forecasting”.

5: Snapshot of moving average analysis Figure 5.snapshot Figure 5. 3) with increasing linear trend Figure 5.2: Summary of the forecasting methods used by Alfa Laval Figure 4.9: Analysis of raw material items Figure 5.4: Delivery from stock (raw material item nr.8: Graph showing forecast and demand in upward trend.1: Raw material item numbers and current forecasting methods Figure 5.Component Unit Plates Mean Absolute Percentage Error Mean Absolute Deviation Mean Squared Error 68 . Forecast made by double exponential smoothing (2exp) Figure 5.2: Major Time Series Models Figure 3.6: Snapshot of double exponential smoothing analysis Figure 5. 01/ 2007 Figure 5.3: Historical demand with corrected values Figure 5.2: Historical demand data .3: Intercept and Slope Figure 3.7:Forecast accuracy.Figure 3.10: Summary of the raw material items and the best forecasting methods List of Abbreviations CU SU OM-CP MAPE MAD MSE Component Unit Supply Unit Operation Manufacturing .1: Current order and forecasting process Figure 4.3: Current forecasting methods for selected raw material items Figure 5.4: Theoretical framework Figure 4.

APPENDIX 1 . Upon what basis and when were the current forecasting models created? 6. Are adjustments documented? 10. What do suppliers think of the prognoses? 11. What are the current forecasting models? How are they weighted? 5. How is demand measured? Using what metrics? 7. How many suppliers are available? 4. Please describe the ordering process for raw materials. Which products are to be analyzed and why? 12. 1. What ERP system do you use? 8. Who is responsible for placing orders? 3. Based on what facts are prognoses created in Jeeves adjusted? 9.Interview questions GENERAL QUESTIONS ABOUT THE ORDERING AND FORECASTING PROCESS Questions from interview with Martin Jönsson and Tobias Augustsson These questions were used throughout the study during unstructured interviews. 2. What are the results of poor forecasting at Alfa Laval? 69 .

Historical data (one item) APPENDIX 3 and 4 . The complete results can be found on the enclosed CD 70 .Raw material analysis (one item) This report contains historical data and raw material analysis for one item only.APPENDIX 2 .

Analysis of evaluating method Analysis of Alfa Laval’s accuracy evaluation method Formula for estimating the forecast error as a percentage (%): et =((D3m – F3m)/F3m)*100 et = forecast error at period t (three months) D3m = Demand (Delivery from Stock) . 14 and 13 which it forecasts.b) / 4) / ((2a +b) / 4) = (2a-b) / (2a+b) It means that the evaluation method is probably incorrect.5*(6+5+4+3+2+1)) / 4 Let us call (12+11+10) for a and 1*(9+8+7) + 0.5*(6+5+4+3+2+1) for b Than we can write Ft3m as (2a+b) / 4 D3m is (12+11+10) = a Formula for the forecast evaluation is: (D3m-Ft3m) / Ft3m Substituting in the formula we obtain: (a -(2a+b) / 4) / ((2a+b) / 4) = ((4a -2a .APPENDIX 5 . There is no indication of how accurate the prediction Ft3m is. 1 2 3 4 5 6 7 8 9 10 D3m 11 e/F3m = (D3m-F3m)/F3m 12 13 Forecast F3m 14 (2a+b)/4 15 b a 71 . 11 and 12). formula is comparing the forecasted future data (next three months 13. it should be compared to recorded real data in months 15. 14 and 15) to the past data (months 10.sum of the past three months F3m = Forecast .usage one year ahead divided by four Forecast for the next three months is: Ft3m = (2*(12+11+10) + 1*(9+8+7) + 0.

72 .