Warehouse Cost Estimation

Tim Bijl
ORTEC-Consulting
August, 2016
Warehouse cost estimation

Date: August 16, 2016

By
Tim Bijl
S1135538
[email protected]

Supervision University of Twente

Department Industrial Engineering and Business Information Systems
Dr. P.C. (Peter) Schuur
Ir. H. (Henk) Kroon

Supervision ORTEC-Consulting
Department Supply Chain Strategy & Excellence
Msc. W. (Wim) Kuijsten
Msc. F. (Frans) v. Helden

Faculty & Educational program

Faculty of Behavioural, Management and Social Sciences
Master: Industrial Engineering and Management
Track: Production and Logistics Management

i
VOCABULARY
Abbreviation Description
ABC Activity based costing
AIMMS Optimization software
Automation level The level of automation within a warehouse
Cost driver An entity that drives/influences the costs
Estimator Independent variable
Extended regression Conceptual regression equation
IPOPT Interior point optimizer, solver for nonlinear optimization problems
MINLP Mixed integer nonlinear problem
OSCD ORTEC Supply Chain Design; a tool designed to perform supply chain
studies at ORTEC
Predictor Dependent variable
RMSE Root mean square error

ii
MANAGEMENT SUMMARY
ORTEC-Consulting helps customers manage their supply chains, by mapping all the locations, flows
and subsequent costs and uses this as input for supply chain studies. This information is all taken into
account in ORTEC Supply Chain Design (OSCD), a tool especially designed for this kind of studies.
Within OSCD the goal is to optimize the design of the supply chain, for different scenarios set by the
user. A possible scenario is adding warehouses to the supply chain of a customer. In order to model
an additional warehouse, it is essential to know how the costs can be determined. At this moment,
no standard procedure is available to assess the periodic costs of a warehouse. Therefore, ORTEC
formulated the following problem statement:

“In order to make good cost estimations for newly built warehouses or depots, build a generic, user-
friendly tool that can quickly and accurately estimate the periodic costs of a warehouse”, with:

- Generic: Regardless of sector and the availability of data, the tool must be able to do accurate
estimations. Basic cost and operational data, such as the total costs and the amount of
products stored in or passing through a warehouse per period, can be expected from all
customers.
- Accurately: Given the situation (the availability of data), the tool must use the most
appropriate method to provide a reasonable estimation. The goal, as set by ORTEC, is to
perform cost estimations with a maximum deviation of 10% in 90% of the cases.
- Tool: The desired platform for this tool is AIMMS. The tool must be designed in such way it
can later be implemented within OSCD.
- Quickly: As part of an OSCD-study the tool must be fast, preferably providing an estimation
within the order of seconds.
- Costs: The total periodic operating costs of a warehouse.

In this research, first the main high-level cost drivers of a warehouse are defined. From literature,
interviews and analysis of a customer-case the following cost drivers are defined:

- Throughput
- Building area
- Labour
- Automation level
- Country and region

of which throughput is presumably the most powerful driver. After having analyzed what kind of data
is provided by the customer for a typical supply chain study, it seems that especially throughput and
the country a warehouse is in are to be expected as input.

Several cost estimation methods have been evaluated mainly based on speed, accuracy and the data
available. After evaluating, several forms of parametric estimation have been selected to apply in a
case-study: simple linear regression, multi-regression, nonlinear regression and a conceptual
extended regression equation. Of these methods, the nonlinear and the simple linear regression are
based on throughput as single estimator for the total costs. The multi-regression method is applied
based on two estimators, namely throughput and the building area. The conceptual extended
regression method was set-up in such a way it can be applied using the country, the area, the
automation level and the throughput of a warehouse as estimators. Since not all these elements were
available for the case-study, only throughput and the country of the warehouse were taken in to
account in the equation.

iii
In addition to the parametric estimation methods, activity based costing is also selected as a cost
estimation method. This method is applied in two ways. The first application is the assignment of all
costs to the throughput, resulting in an average cost per unit. The second application is similar, but
now the average cost per unit is defined per country.

All the cost estimation methods have been implemented in AIMMS and the optimization engine was
used to define the slope and intercept of the different regression methods. The clustering of
observations is also done by optimizing a mathematical model. The analysis, as well as the actual cost
estimation is all built-in in AIMMS.

The best performing method is nonlinear regression, with throughput as independent variable and
an exponent of approximately 0.7. Other strong results are reported for extended regression, based
on throughput and the country, and multi-regression based on both throughput and building area.

This high-level approach resulted in meeting all the requirements, except for accuracy. The nonlinear
model reports to estimate the costs of a warehouse within 10% deviation in 39% of the cases. 90%
estimation accuracy is only acquired if 40% deviation is allowed (96%). Therefore, the following
recommendations are formulated:

- Gather datasets from customers with more cost drivers available and preferably more
observations. In this way, the analysis can be more thorough and more sophisticated cost
estimation models can be developed and evaluated. This will likely increase the accuracy of
the estimations.
- Collect multiple datasets of different customers within different industries. This research is
based on one individual client, so it would be interesting to see if the same conclusions can
be drawn over a combined dataset containing multiple clients and industries. In this way
general rules can be developed. Other methods, like machine learning could also be applicable
to these larger datasets.

Further research is required for the following topics:

- More datasets with more cost drivers must be collected in order to develop more accurate
cost estimation models.
- Combined datasets from different customers and sectors could provide general rules that are
applicable to every customer in every sector.
- A distinction can be made between several cost entries and the factors that drive these.
Investigating this may lead to more accurate cost estimations.
- Find a good balance between the gathering of data and the power of the estimations.

iv
PREFACE
After having many great years as a student at the University of Twente, I hereby present my master
thesis. With this thesis I complete my master program Industrial Engineering and Management with
as specialization Production & Logistics Management. The research described in this report is executed
at ORTEC-Consulting in Zoetermeer, where I also worked part-time as student assistant.

I would like to thank all the people I enjoyed my time with in Enschede: my roommates at
‘Studentenhuis Fortes’, my sorority Pineut and all other people I met and spent time with.

Furthermore, I would like to thank my supervisors at ORTEC, Wim and Frans, and all other people who
helped me and provided feedback or relevant insights. I appreciate the time Wim spent supervising
me and the freedom he gave to do my research.

I enjoyed my time at ORTEC very much and I also learned a lot from the people of my team. I really
liked working with the people at ORTEC, during my thesis and other projects I was involved in.

I would also like to thank Peter Schuur, my supervisor of the University of Twente, for his contribution
to my research and his positive way to look at things. Also thanks to my second supervisor, Henk
Kroon, for his support and critical reviews.

I would like to thank all the people who made time to support me or help me by reviewing my thesis
report.

A special thanks to my parents, who supported me my entire (extensive) study in both Amsterdam
and Enschede.

Den Haag, October 1st, 2016

Tim Bijl

v
 TABLE OF CONTENTS
Vocabulary .............................................................................................................................................. ii
Management summary.......................................................................................................................... iii
Preface .................................................................................................................................................... v
Table of Contents ................................................................................................................................... vi
List of figures .......................................................................................................................................... ix
List of tables ........................................................................................................................................... ix
1 Introduction and problem description ........................................................................................... 1
1.1 ORTEC & ORTEC-Consulting .................................................................................................... 1
1.2 Business context ..................................................................................................................... 1
1.3 Problem description ................................................................................................................ 2
1.4 Problem statement & scope ................................................................................................... 2
1.5 Research questions ................................................................................................................. 2
1.6 Deliverables............................................................................................................................. 3
1.7 Running example .................................................................................................................... 3
1.7.1 Customer characteristics................................................................................................. 4
1.7.2 Business request ............................................................................................................. 4
1.8 Report outline ......................................................................................................................... 5
2 Relevant cost drivers ....................................................................................................................... 7
2.1 Typical warehouse costs ......................................................................................................... 7
2.2 Cost drivers from literature .................................................................................................... 8
2.3 Cost drivers from interviews ................................................................................................... 9
2.4 Cost drivers from case study ................................................................................................. 10
2.4.1 Raw data ....................................................................................................................... 10
2.4.2 Normalized data ............................................................................................................ 11
2.4.3 Clustering ...................................................................................................................... 12
2.5 Conclusion ............................................................................................................................. 14
3 Data availability............................................................................................................................. 15
3.1 Typical input for supply chain studies ................................................................................... 15
3.2 Data request.......................................................................................................................... 15
3.3 Data provided by customers ................................................................................................. 16
3.4 Conclusion ............................................................................................................................. 16
4 Cost estimation methods .............................................................................................................. 17
4.1 Overview methods ................................................................................................................ 17
4.2 Parametric estimating ........................................................................................................... 18

vi
 4.2.1 Simple linear regression ................................................................................................ 18
4.2.2 Multiple linear regression ............................................................................................. 20
4.2.3 Nonlinear regression ..................................................................................................... 21
4.2.4 Clustering ...................................................................................................................... 22
4.2.5 Machine learning .......................................................................................................... 23
4.2.6 Activity based costing ................................................................................................... 24
4.3 Engineering build-up ............................................................................................................. 25
4.3.1 CCET .............................................................................................................................. 25
4.4 Analogy ................................................................................................................................. 25
4.5 Expert opinion ....................................................................................................................... 25
4.6 Summary cost estimation methods ...................................................................................... 26
4.7 Method comparison and choice ........................................................................................... 26
4.8 Conclusion ............................................................................................................................. 28
5 Approach & mathematical formulation ........................................................................................ 29
5.1 Simple linear regression & multi-regression......................................................................... 29
5.1.1 Sets and indices............................................................................................................. 29
5.1.2 Parameters .................................................................................................................... 29
5.1.3 Variables........................................................................................................................ 29
5.1.4 Problem statement ....................................................................................................... 29
5.1.5 Problem classification ................................................................................................... 30
5.2 Extended regression ............................................................................................................. 30
5.2.1 Sets and indices............................................................................................................. 31
5.2.2 Parameters .................................................................................................................... 31
5.2.3 Variables........................................................................................................................ 31
5.2.4 Problem statement ....................................................................................................... 31
5.2.5 Problem classification ................................................................................................... 32
5.3 Nonlinear regression ............................................................................................................. 32
5.3.1 Sets and indices............................................................................................................. 32
5.3.2 Parameters .................................................................................................................... 32
5.3.3 Variables........................................................................................................................ 32
5.3.4 Problem statement ....................................................................................................... 32
5.3.5 Problem classification ................................................................................................... 32
5.4 Clustering .............................................................................................................................. 32
5.4.1 Sets and indices............................................................................................................. 33
5.4.2 Parameters .................................................................................................................... 33
5.4.3 Variables........................................................................................................................ 33

vii
 5.4.4 Problem statement ....................................................................................................... 33
5.4.5 Problem classification ................................................................................................... 33
5.5 ABC ........................................................................................................................................ 33
5.5.1 Sets and indices............................................................................................................. 33
5.5.2 Parameters .................................................................................................................... 34
5.5.3 Problem statement ....................................................................................................... 34
5.5.4 Problem classification ................................................................................................... 34
5.6 Conclusion ............................................................................................................................. 34
6 Case study ..................................................................................................................................... 35
6.1 Set-up case study .................................................................................................................. 35
6.2 Testing procedure ................................................................................................................. 35
6.3 Results ................................................................................................................................... 36
6.4 Results with respect to goal ORTEC ...................................................................................... 37
6.5 Conclusion ............................................................................................................................. 38
7 Implementation ............................................................................................................................ 39
7.1 Implementation .................................................................................................................... 39
7.2 User interface........................................................................................................................ 39
7.3 Performance ......................................................................................................................... 40
7.4 Conclusion ............................................................................................................................. 40
8 Conclusions & Discussion .............................................................................................................. 41
8.1 Conclusion ............................................................................................................................. 41
8.2 Recommendations ................................................................................................................ 43
8.3 Discussion.............................................................................................................................. 43
8.4 Future research ..................................................................................................................... 44
Bibliography .......................................................................................................................................... 45
Appendix A: Opening page ................................................................................................................... 47
Appendix B: Clustering .......................................................................................................................... 47
Appendix C: Estimation page ................................................................................................................ 48

viii
LIST OF FIGURES
Figure 1.2-1: Visualization of OSCD......................................................................................................... 1
Figure 1.7-1: Current (blue) and proposed (green) locations of southern warehouses of the packaging
distributor ............................................................................................................................................... 4
Figure 2.1-1: Simple warehouse cost-tree (from Richards, 2010) .......................................................... 8
Figure 2.2-1: Stepwise linear warehousing cost function (From Goh et al. 2001) ................................. 8
Figure 2.3-1: Inter-relationships of cost drivers based on interviews .................................................. 10
Figure 2.4-1: Scatterplot of throughput versus total costs (implemented in AIMMS) ......................... 10
Figure 2.4-2: Scatterplot of building area versus total costs (implemented in AIMMS) ...................... 11
Figure 2.4-3: Number of clusters versus explained variance for Customer 2....................................... 12
Figure 2.4-4: Division of the countries of origin of the observations over three clusters .................... 13
Figure 4.2-1: Regression line of building area versus the total costs (implemented in AIMMS) ......... 19
Figure 4.2-2: Nonlinear regression line of throughput versus the total costs (implemented in AIMMS)
.............................................................................................................................................................. 21
Figure 4.2-3: Clustering procedure (from Xu, 2005) ............................................................................. 22
Figure 4.2-4: Traditional data analysis versus algorithmic modeling (From Breiman, 2001) ............... 23
Figure 4.2-5:Example of how a decision tree for cost estimation may look like .................................. 24
Figure 4.7-1: Performance matrix of different estimation methods .................................................... 28
Figure 5.2-1: Cost drivers of labour costs ............................................................................................. 30
Figure 5.2-2: Cost drivers of building area costs ................................................................................... 30
Figure 6.4-1: Accuracy per method....................................................................................................... 38

LIST OF TABLES
Table 2.4-1: Strongest results found within the provided data-set...................................................... 11
Table 2.4-2: Strongest results found within the provided data-set...................................................... 11
Table 2.4-3: Strongest results found within the provided normalized data-sets ................................. 12
Table 2.4-4: Average values per cluster, with 2 clusters ...................................................................... 13
Table 2.4-5: Average values per cluster, with 3 clusters ...................................................................... 13
Table 3.3-1: Elements available within data-sets analyzed .................................................................. 16
Table 4.1-1: Three cost estimation methods compared (from Leonard, 2009) ................................... 17
Table 4.2-1: Regression equation based on building area .................................................................... 20
Table 4.2-2: Regression equation based on throughput and building area ......................................... 20
Table 4.2-3: Regression equation based on throughput ...................................................................... 21
Table 4.4-1: Example of the analogy cost estimating method (from Leonard, 2009) .......................... 25
Table 4.7-1: Different estimation methods with their score per criterion ........................................... 27
Table 6.1-1: Independent variables present within the dataset of Customer 2................................... 35
Table 6.2-1: Estimation methods and used predictors ......................................................................... 35
Table 6.2-2: Number of observations per country used for the test .................................................... 36
Table 6.3-1: Results of the cross-validation, expressed in RMSE ......................................................... 36
Table 6.3-2: Average equation of the three most accurate cost estimation methods ........................ 37

ix
x
1 INTRODUCTION AND PROBLEM DESCRIPTION
In the context of my Master’s thesis I performed this research at ORTEC-Consulting in Zoetermeer,
thereby finishing the master track Production & Logistics Management at the University of Twente.

In this first chapter, the company ORTEC is described as well as the need for this research.
Furthermore, the main goal as well as the research questions that result in achieving the research goal
are set out.

1.1 ORTEC & ORTEC-CONSULTING

ORTEC is founded in 1981 by five econometrics students from the Erasmus University in Rotterdam.
The founders believed the mathematical theories and algorithms they worked on could be practically
applied to improve business performance. Today, ORTEC is serving clients in almost every industry
and has 15 offices located across 4 continents with around 900 employees, of which most are highly
educated with a quantitative background. The ORTEC headquarters is situated in Zoetermeer, a city
in the west of the Netherlands, and hosts about 400 employees.

ORTEC-Consulting is one of the three components that form ORTEC, next to Products and Living Data.
The headquarters in Zoetermeer hosts about 130 consulting employees and is active in many fields.
Large customers are in oil & gas, aviation and consumer packaged goods and the gross of the projects
is on a tactical or strategical level. Almost all solutions provided by ORTEC-Consulting are
quantitatively based, which distinguishes ORTEC from its competition.

1.2 BUSINESS CONTEXT

Supply chain studies are one of the main activities within ORTEC Consulting. Within these studies,
customers request insight in their supply chain and strategic (semi-) optimal choices are to be made
with respect to production and stock levels, routing, opening and/or closing locations. For this
purpose, ORTEC designed its own tool, ORTEC Supply Chain Design (OSCD), to perform these kind of
studies. An example of how the outcome of an OSCD study can be visualized can be seen in Figure 1.2-
1. OSCD is implemented in AIMMS, a software system designed for modeling and solving large-scale
optimization and scheduling-type problems, in which these kind of optimization problems can easily
be modelled.

Figure 1.2-1: Visualization of OSCD

1
1.3 PROBLEM DESCRIPTION
The focus of this research is at opening and/or closing warehouses, a part of the study normally carried
out with Greenfield analysis, also integrated in OSCD. In these kind of studies, it is essential to gather
as much information as possible. Supply, demand, capacities, coordinates and costs are all taken into
account in balance equations within the mathematical model which will be eventually optimized.
Different scenarios are set up, including choices about closing and opening different warehouse
locations, in order to evaluate the financial consequences of different strategies. An important part
within these studies is estimating the periodic operational costs of a new warehouse, because it has a
major impact on the outcome of the different scenarios. If the estimation is poor and not well-
founded, the quality of the solution is likely to be poor as well.

In the current situation the costs are in many cases estimated by doing a side-study for a given set of
potential locations, based on characteristics of already existing warehouses and their cost-structure.
There is no standard procedure for this cost estimation and it is case-dependent which approach is
chosen for. Fact is that the current approach is time-consuming, not standardized and performance is
not guaranteed.

1.4 PROBLEM STATEMENT & SCOPE

To fill the gap between the current and the desired situation, this research aims for providing an
accurate cost estimation method for warehouses or depots. Therefore, the research statement and
the scope of the research can be summarized as follows:

“In order to make good cost estimations for newly built warehouses or depots, build a generic, user-
friendly tool that can quickly and accurately estimate the periodic costs of a warehouse”, with:

- Generic: Regardless of sector and the availability of data, the tool must be able to do accurate
estimations. Basic cost and operational data, such as the total costs and the amount of
products stored in or passing through a warehouse per period, can be expected from all
customers.
- Accurately: Given the situation (the availability of data), the tool must use the most
appropriate method to provide a reasonable estimation. The goal, as set by ORTEC, is to
perform cost estimations with a maximum deviation of 10% in 90% of the cases.
- Tool: The desired platform for this tool is AIMMS. The tool must be designed in such way it
can later be implemented within OSCD.
- Quickly: As part of an OSCD-study the tool must be fast, preferably providing an estimation
within the order of seconds.
- Costs: The total periodic operating costs of a warehouse.

1.5 RESEARCH QUESTIONS

Following the problem statement, several research questions must be answered in order to come to
a solid solution. As the requirements and specifications are already covered within the problem
statement, these do not need to be addressed separately. However, an important limitation from the
problem statement is the availability of data, which leads to the first research question:

1. What kind of data can be expected from the customer?

This question will be answered by looking at data of past projects and by requesting specific data from
a current customer.

2
 2. What are the most important cost drivers to be identified?

Analysis of data, as provided by customers, scientific insights and expert opinions are used to form
some hypothesis about cost drivers and relations. These hypotheses will be tested later in this
research.

3. What relevant methods are available to do cost estimations?

After determining the limitations and requirements, several methods from scientific literature will be
evaluated and finally some choices will be made, leading to the following research question:

4. Which method or methods are most suitable given the situation at ORTEC?

The several methods will be evaluated carefully, leading to the choice of a method that best suits the
purpose of this research and taking into account the restrictions and limitations.

5. How can the insights gained in this research be implemented?

The results and insights following from this research must be implemented within AIMMS, an
optimization platform. The design choices, based on the requirements set, are therefore discussed
and explained.

Throughout this thesis, all research questions are answered and the insights gained from this form the
basis for solving the problem statement. At the end of this report a conclusion is drawn about the
resulting method or methods and whether it meets all the requirements.

1.6 DELIVERABLES
This research will be partly based on insights from literature and experts and partly based on a case
study. Since not much research is done in the field of warehouse cost estimation, this research is
heavily based on the insights gained from provided data. The insights gained can be of particularly
good use for ORTEC or any other company that aims at reviewing or investing in their supply chain.

To summarize, this research results in providing the following deliverables:

- Define what kind of data can be expected from customers.

- Provide insight in what factors have the highest influence on the costs of a warehouse. This
means that the most important cost drivers will be identified and later be used as the basis of
cost estimation techniques.
- Provide an overview of available cost estimation methods.
- Determine the best cost estimation technique for warehouses.
- Apply the gained insights by developing a tool in AIMMS that can analyze data and provide
estimations.

1.7 RUNNING EXAMPLE

In order to underline the need for this research, as well as to illustrate several described methods, a
running example is introduced. The running example is based on an actual customer of ORTEC-
Consulting and its data is used for the case study. In the following sections, the customer
characteristics are described as well as a (fictional) business request. In the remainder of this thesis,
this customer will be referred to as the Packaging distributor.

3
1.7.1 Customer characteristics
The packaging distributor is a provider of reusable packaging in the European fresh supply chain. In
order to efficiently distribute the large amount of homogeneous products throughout Europe, about
100 warehouses are spread over 14 different countries in Europe. Within these warehouses, the
products are moved in, sorted and washed before getting stocked. When stocked, the products are
ready to be distributed to its customers. The amount of products stocked is dependent on seasonal
demand; typically, in periods with high demand, the stock-levels will be low. On average, the amount
of products annually going through a warehouse is about 30 million. The processes inside the
warehouse are not labor-intensive, and the employees responsible for the internal handling are mainly
temporary workers.

1.7.2 Business request

The business of the packaging distributor is growing. Especially in the south of Europe: in Spain, France
and Italy there is growing demand for the packaging, such that the current southern warehouses
cannot meet the demand. Therefore, products have to come from other warehouses, further away,
which means that the transport costs rise and profit decreases.

In order to cope with this changing situation, the packaging distributor wants to review his supply
chain and thinks about expanding its warehouse capacity in the south of Europe. Since the capacity of
the current warehouses is not further expandable, the capacity can only be extended by acquiring new
warehouse space. The packaging distributor already has two potential locations in mind, as illustrated
in Figure 1.7-1.

1. Toulouse, France:
- Urban area in the south of France
- Close to important highways in France and to Spain

2. Badojz, Spain
- Rural area in the west of Spain, next to the border with Portugal
- Close to a main road to Portugal

Figure 1.7-1: Current (blue) and proposed (green) locations of southern warehouses of the packaging distributor

The packaging distributor wants ORTEC to find the optimal design of its supply chain, which is in this
case the most profitable. In order for ORTEC to calculate optimal solutions for different scenarios, it
needs operational and cost data of all locations. In addition, it needs an estimation of the operational

4
costs of the two proposed warehouses. With this data ORTEC can calculate the effects of closing and
opening different locations and give the packaging distributor a well-founded advice.

It is likely to be the case that the proposed warehouse locations will have different cost structures,
since:

- They are in different countries: different labor-rates;

- One is close to a city, while the other lies in a more rural area: different land costs;
- They presumably differ in size, capacity and automation level: effects all operational costs;

This thesis provides in identifying the most important cost drivers and uses these to accurately
estimate the costs of warehouses. These estimations can be used in the kind of studies as described
in the example above. Because this example is mainly meant for illustrating the need for accurate
warehouse cost estimation and also forms the basis for some examples throughout this thesis, there
is no need for further details regarding the supply chain study.

1.8 REPORT OUTLINE

This report provides answers to the research questions, as stated in the section 1.5, and eventually
the problem statement. First the data availability is determined, thereby answering the first research
question. The most important cost drivers are studied in chapter 3, based on literature, interviews and
a case study. In chapter 4, relevant estimation methods are described and the most suitable
approaches are selected based on the requirements and scope of this research.

In chapter 6 the approach and set-up of the case study is described, together with the mathematical
formulation. The results of this case study are presented in the same chapter. In chapter 7 a
description of the implementation in AIMMS is given and the last chapter provides conclusions and
further recommendations.

5
6
2 RELEVANT COST DRIVERS
After describing the goal and defining the purpose of this research in Chapter 1, this chapter aims at
identifying the factors that drive the costs of a warehouse.

In order to answer research question 2:

2. What are the most important cost drivers to be identified?

the following steps are taken:

- Presenting a global overview of the typical costs a warehouse is facing

- Taking into account expert-opinions, to determine what factors emerge from practice
- Conducting a case study, to determine which cost drivers are found through data-analysis1
- Discussing relevant insights from scientific literature

The results from each section are taken into account and a conclusion is drawn about the most
important cost drivers. The insights gained from this chapter will be used as basis for the warehouse
cost estimation and the relative impact of each of the factors will be evaluated later in this research.

2.1 TYPICAL WAREHOUSE COSTS

Before determining what factors drive the costs of a warehouse, it is essential to define these costs.
In order to get a clear view of what typical expenses a warehouse faces and which take the most part
of it, a cost-tree is helpful. Such a cost-tree is defined by Richards (2010), and is presented in Figure
2.1-1. Richards (2010) distinguishes three major cost components, namely storage, handling and
overhead costs, which eventually can all be broken down into direct expenses. In addition to having a
clear overview of the different costs a warehouse faces, the cost-tree reflects different levels of detail.
The higher the level of detail, the more accurate an estimation will be and the more time it costs to
perform such an estimation.

Furthermore, Figure 2.1-1 shows the percentage per element of which labour, with 60 percent, takes
up the most part of the storage and handling costs. Followed by space and equipment costs with
respectively 25 and 15 percent of the total storage and handling costs.

The cost-tree of Richards (2010), does not include the proportion of overhead in relation to the total
costs. According to several experts within ORTEC, the most common variable/fixed costs ratio is 60/40,
but the exact ratio is highly dependent on the sector and customer that is dealt with.

Summarizing this, the biggest costs components are:

- Labour
- Space & equipment
- Overhead

In the remaining of this chapter the cost drivers of the main expenses, as discussed above, are
determined by conducting scientific literature, interviews (expert-opinion) and by analyzing cases.

1
The data-analysis is mainly based on methodology as described in chapters 4 and 6.

7
Figure 2.1-1: Simple warehouse cost-tree (from Richards, 2010)

2.2 COST DRIVERS FROM LITERATURE

Several studies have been performed in the field of warehousing, with different purposes. Goh et al.
(2001), conducted a study about warehouse sizing with as main goal minimizing inventory and storage
costs. Goh et al. solve this problem by modelling the warehouse costs as a piecewise linear function,
which implies that different sized warehouses have different cost-structures and that the size (or
building area) of a warehouse is an important cost driver. In the model of Goh et al. size and
throughput are the main cost drivers, as the variable and the fixed costs are assumed given and the
total variable costs are driven by throughput (see Figure 2.2-1).

Figure 2.2-1: Stepwise linear warehousing cost function (From Goh et al. 2001)

Hung & Fisk (1984) present a linear programming model to determine the amount of warehousing
space a firm should buy when confronted with highly seasonal demand. They model the variable
warehousing cost for a certain period using the cost per square feet and the demand for storage space.
The overhead costs are determined using the overhead per square feet and the size of the warehouse.
The costs in this model are driven by throughput (demand for storage space), the space utilization of
the products, the country or region (the cost per square feet) and the size of the warehouse.

8
Young & Webster (1989) developed an optimization procedure to help a warehouse planner in the
design of selected three-dimensional, palletized storage systems. The linear programming formulation
is based on an extensive cost model, that is used to find a design with minimal costs. The model
includes the following periodic expenses:

- Land costs: Modelled by multiplying the total needed building area with the costs per square
feet
- Building costs: Modelled by using the equipment costs per square feet and height, a utilization
factor and the total needed building area
- Equipment costs: First the required number of pieces of equipment is determined using the
preferred handling rate and internal travel time of products. The total equipment cost is then
acquired taking the sum of the pieces of equipment plus the affiliated conveyor system cost
and the control system cost.
- Storage rack facility costs: These costs are based on the kind of storage racks and the amount
of racks needed.
- Labour costs: The authors base the number of employees needed on the amount of people
needed per piece of equipment (so not on the actual throughput). In their formulation this
costs get corrected for inflation.
- Maintenance costs: The maintenance costs consist of maintenance of the building, the control
system and the equipment, where the building maintenance is dependent on square feet and
the equipment maintenance depends on the equipment choices.
- Operating costs: The operating costs consist of charging batteries and fuel of the handling
equipment.

This detailed way of modelling the warehouse costs is beyond the scope of this research, but the
important cost drivers can be derived from this model. The land-, building- and maintenance-costs are
mainly driven by the total needed building area for the warehouse and the country (or region)
specifics, such as land costs or maintenance costs. The equipment-, operating- and labour-costs are
all driven by the kind of product and the amount of it, since the choice of equipment and the number
of employees is based on these factors. Next to that, the country (or region) is a cost driver here, as it
influences the wages.

Although differently modelled, the cost formulations described in this section are all driven by
throughput and the size of the warehouse. Additional cost drivers are the country or region the
warehouse is located and the product type, since space utilization as well as the choice of equipment
is driven by product characteristics.

2.3 COST DRIVERS FROM INTERVIEWS

Based on interviews with experts in the field of supply chain studies or warehousing in general, within
ORTEC-Consulting, some presumptions can be made about cost drivers. The majority of the
interviewees stated that the physical location of the warehouse is of major influence on the periodic
costs, due to fluctuation of employee and land costs. A distinction must be made between the country
where the warehouse is located and whether the actual location is in a rural or an urban area. Other
important cost drivers mentioned are throughput (or capacity), building area in square meters, the
number of employees and the automation level.

None of these mentioned cost drivers is totally independent and co-relations exist. Several
interviewees stated that the location, that is country and/or area, together with the automation level
and the throughput are the main cost drivers. The number of employees and the building area of the

9
location follow directly from the other cost drivers. When translated into a logical model, this looks as
described in Figure 2.3-1. The mentioned cost drivers are marked blue, the resulting costs are marked
green and additional factors are white. To reduce the amount of overlapping arrows, the lower cost
drivers are grouped. Overhead costs are not taken into account in this model, since these are sector-
and company-specific.

Figure 2.3-1: Inter-relationships of cost drivers based on interviews

2.4 COST DRIVERS FROM CASE STUDY

The goal of this section is to test the cost drivers derived from literature and experts and determine
which factors are putting the most weight into the equation. Therefore, the first step is to examine
the raw data out of the data of the packaging distributor using linear regression to see the relation
between the proposed cost drivers and the total costs, as well as relations between estimators. To
check whether a stronger fit is found when normalizing for labour costs, the same tests are performed
using normalized data. To further investigate possible underlying relations, between countries for
example, the data is split using a clustering algorithm. The implementation of both simple linear
regression and clustering is explained in Chapter 4.

2.4.1 Raw data

When starting an analysis, it is essential to first explore the data visually. Therefore, two scatterplots
are presented, in Figure 2.4-1 and 2.4-2. indicating the relation between throughput and the total
costs and building area and the total costs. The outliers in the data are detected based on standard
deviation. The data is fitted with a linear regression line and observations that lie outside this line
plus standard deviation are removed from the analysis.

Figure 2.4-1: Scatterplot of throughput versus total costs (implemented in AIMMS)

10
Figure 2.4-2: Scatterplot of building area versus total costs (implemented in AIMMS)

As to be seen from the scatterplots, the spread of the total costs, the throughput and the building
area is large. It is also clear that both throughput and building area behave quite linear towards the
total costs.

To put these observations into a more quantitative view, linear regression is applied to the data, of
which the results are presented in Table 2.4-1. There seems to be a strong relationship between
throughput and the total costs and building area and the total costs. Multiple regression shows an
even stronger relationship where both the building area as the throughput are taken into account in
the regression equation.
Table 2.4-1: Strongest results found within the provided data-set

Observations Estimator Predictor R2 (adj.) Correlation

69 Throughput Total costs 0.90 0.95
42 Building area Total costs 0.85 0.92
34 Building area + Total costs 0.96 -
Throughput

To check what influence the throughput has on the building area, the R2 of throughput as estimator
and building area as dependent variable is also calculated for Customer 2. The results can be found in
Table 2.4-2. Although the R2 is not as high as with total costs as predictor, throughput does seem to
have influence on the building area of a warehouse, which indicates that throughput and building area
are far from independent.
Table 2.4-2: Strongest results found within the provided data-set

Observations Estimator Predictor R2 (adj.) Correlation

26 Throughput Building .76 0.87
area

2.4.2 Normalized data

The dataset contains observations from different countries in Europe, where large wage differences
may be expected. To investigate if this can be corrected for, using labor cost index from Eurostat

11
(2015), the total costs are normalized using this cost index. The same estimators and the same
technique as for the basic data are applied to this normalized data, to see whether it improves the fit
within the data.

As to be seen in Table 2.4-3, the fit within the normalized data is worse than for the initial data, for all
set-ups tested. Normalizing data for wage differences does not seem to be an appropriate method to
increase the fit or the estimating power. The country and location influence cannot be totally ruled
out, because other factors besides wage may be of influence, and the differences might be specific
per company.
Table 2.4-3: Strongest results found within the provided normalized data-sets

Observations Estimator Predictor R2 (adj.) Correlation Old R2

(adj.)
67 Throughput Total Costs 0.87 0.93 0.90
40 Building area Total Costs 0.41 0.64 0.85
32 Building Total Costs 0.89 - 0.96
area+
Throughput

2.4.3 Clustering
To further investigate underlying relations within the data, a cluster algorithm is used to divide the
data into subsets resulting in a higher fit within the cluster. After clustering, the clusters are compared
to determine on which characteristics they differ and see whether this provides additional insight
regarding cost drivers.

The cluster algorithm is implemented using a mixed-integer nonlinear program (MINLP) -formulation,
minimizing the sum of the squared error over all the clusters, following the cluster-wise linear
regression heuristic of Späth (1978). This model is rewritten to an exact optimization algorithm, of
which the mathematical model can be found in Chapter 5. The implementation of this algorithm is
done by relaxing the binary variable, to improve solving speed, resulting in an acceptable
approximation of the global optimum.

With the number of clusters increasing, the squared error drops, leading the percentage of explained
variance to rise to 100 percent. Common practice is to identify the ‘elbow’ (Thorndike, 1953), the point
after which the marginal gain of the explained variance drops. As to be seen in Figure 2.4-3, in which
clustering is applied to the data of the packaging distributor, the marginal gain in explained variance
is the highest when shifting from two to three clusters. Since the average dataset will not exceed 40
observations, more than three clusters will generally not add additional insights. Two or three clusters
will generally be sufficient.

Figure 2.4-3: Number of clusters versus explained variance for Customer 2

12
The cluster analysis is performed on the data of the packaging distributor. Initially, the method is used
to analyze the characteristics of the formed clusters. Clustering is also tested as estimation method,
in Chapter 6.

For the first experiment, the observations are clustered in two clusters based on throughput as
estimator versus the total costs as predictor. The results are presented in Table 2.4-4. It may be
expected that the throughput as well as the total costs differ for each cluster, since the focus of the
clustering was based on these factors. This seems to be the case, and in addition the building area and
automation level seem to scale with the throughput2.
Table 2.4-4: Average values per cluster, with 2 clusters

Cluster- Cluster # Warehouse Avg. Avg. Avg. Avg. R2 Old

base Observations Throughput Building Automation Costs R2
area level (1-5)
Throughput 1 50 15,774,742 78,904 1 597,149 0.94 0.90
2 19 73,262,283 201,644 3 4,069,350 0.90 0.90
For the second experiment, the observations are divided into three clusters based on throughput as
estimator variable versus the total costs as predictor. The results are presented in Table 2.4-5. The
same trends are visible: clear clusters when it comes to throughput and total costs and building area
and automation level that scale with the throughput.
Table 2.4-5: Average values per cluster, with 3 clusters

Cluster- Cluster # Warehouse Avg. Avg. Avg. Avg. R2 Old R2

base Observations Throughput Building Automation Costs
area level (1-5)
Throughput 1 43 9,743,548 61,320 1 303,182 0.96 0.90
2 10 67,360,195 224,507 3 4,323,100 0.94 0.90
3 16 68,009,124 159,932 3 3,125,454 0.98 0.90
When looking at the country of origin of the observations within each cluster, these are not strictly
divided but mainly spread over the clusters, as shown in Figure 2.4-4. It can, however, easily be seen
that some countries are overly represented in clusters. For example, German and Italian warehouses
are mainly placed in cluster 1, the cheapest warehouses. Therefore, the influence of the country a
warehouse is located is not directly clear but cannot be ruled out either.

Figure 2.4-4: Division of the countries of origin of the observations over three clusters

2
Not all observations contained information about the automation level and the surface, therefore the average
of the observations that do contain this is used

13
2.5 CONCLUSION
In this chapter, the following research question is answered:

2. What are the most important cost drivers to be identified?

Therefore, first the major cost components of a warehouse are defined, based on a cost-tree defined
by Richards (2010):

- Labour
- Space & equipment
- Overhead

In order to evaluate which cost drivers have the highest impact on these cost components, relevant
insights from interviews at ORTEC are stated. The insights of the interviewees are based on experience
with supply chain studies or knowledge of warehouses in general. Next to interviews, scientific
literature is consulted and a case study is performed to identify and test different cost drivers.

The most important cost drivers identified are:

- Throughput
- Building area
- Labour
- Automation level
- Country and region

of which labour can presumably be estimated based on throughput and automation level (and the
country and region).

The impact of these cost drivers is tested on the data of the packaging distributor, the running example
in this research. Not all factors were available, so only throughput, building area and the country are
taken into account. From the data analyzed it seems that both throughput and building area are strong
estimators of the total costs, based on simple regression analysis. Applying multi-regression, with both
throughput and building area as estimator results in an even higher fit with the total costs.

The location factor is attempted to take into account by normalizing the total costs for the European
labour rate per country. The same analysis as for the basic data is applied, but the estimation power
of the cost drivers decreases.

In order to gather additional insights or underline earlier found relations, a clustering algorithm is
applied which provided strong results. The automation level and size of the warehouse seem to scale
with the throughput (Table 2.4-3). This can indicate that both factors do not need to be used as
independent variable, because they are explained by the throughput. Furthermore, the influence of
the country of the origin on the total costs cannot be ruled out according to the results in Figure 2.4-
4.

In the next chapter is determined whether customers are able to supply all the needed cost drivers,
and if not, what kind of data is to be expected.

14
3 DATA AVAILABILITY
Now that the most important cost drivers of a warehouse are determined, the data provided by
customers is examined. This means that in this chapter the main limitation of this research is set out:
the data availability. In order to make good estimations it is essential to know what kind of data may
be expected, as well as the level of detail within this data, thereby answering research question 1:

3. What kind of data can be expected from the customer?

First, the typical data need for supply chain studies is described, followed by a data request at the
packaging distributor. After that, the obtained data of the packaging distributor and two other
customers is set out, to get an impression of the kind of data that is usually provided. At the end of
this chapter, a conclusion is drawn.

3.1 TYPICAL INPUT FOR SUPPLY CHAIN STUDIES

Customers usually record tons of data, but it is highly dependent on the customer whether it has the
needed data available in the right amount. The coordinates, the storage capacity per product including
expansion possibilities of the facility, and the associated costs for moving and storing products are
typical input for supply chain studies, according to interviewees at ORTEC. Using this data, optimal
throughput per location as well as optimal routes and quantities shipped between facilities can be
obtained, in order to optimize the objective (e.g. minimizing costs).

3.2 DATA REQUEST

After identifying the most important cost drivers in Chapter 2, a data-request is made to the packaging
distributor based on these findings. The data-request looked as follows:

Data-request:

All data is per given time-period (e.g. annually/monthly)

- Definition of processes per warehouse

o With per process:
- Throughput
- Dwell time per product
- FTE
- Throughput per FTE (capacity employees)
- Building area needed/product [m2]
- Building area of total process [m2]
- Operating costs
- Overhead entries
o With per entry:
- Total costs
- Estimation of fixed percentage of the costs
- General characteristics
- Average wage of operational staff
- Average wage of management staff
- Country/region
- Land costs per m2
- Total building area of warehouse

15
In case the high-level cost drivers, as defined in Chapter 2, do not provide sufficient estimations, more
detailed data is requested as well, to perform further analysis if needed. The data provided by the
packaging distributor based on this request, as well as two other datasets are discussed in the
following section.

3.3 DATA PROVIDED BY CUSTOMERS

Typically, customers do not provide data in a structured and detailed way, such that it would be ideal
to analyze it. In many cases the data as provided is unstructured, missing values or the desired level
of detail is not present. To explore the kind of data that is expected to be provided by customers for
the cost estimation study, a few cases are analyzed. The datasets analyzed are used in actual supply
chain studies and give an indication of the kind of data that is provided by customers.

I. The first dataset is that of the packaging distributor, the running example.
II. The second dataset looked at is that of an international courier delivery services
company from the Netherlands, which data is actually used to perform a cost estimation (by
hand) as part of a supply chain study.
III. The third dataset is that of an industrial service provider, offering mechanical engineering
components and associated technical and logistical services.
Table 3.3-1: Elements available within data-sets analyzed

Dataset #Warehouses Buildin #Employees Automation Throughput Country Location Total

observed g area level costs
1 87 x x x x
2 32 x x x x x
3 80 x x

The first dataset, of the packaging distributor, contained the richest data. Although the data-request
was much more detailed, the provided data only consists of high-level parameters. The other two
datasets contain even less useful data. Dataset 2 has many holes in the data, which results in less
useful observations. Next to that, the data is quite messy and seems unreliable. Dataset 3 does contain
even less data. It does only contain the amount of orders, the total costs and the location of the
warehouse. Since the amount of orders does not seem to be of much use, because the size of these
orders is not known, it is not taken into account in table 3.3-1.

3.4 CONCLUSION
In this chapter, the following research question is answered:

2. What kind of data can be expected from the customer?

After observing three real-life cases from actual customers, it becomes clear that customers all provide
different kinds of data with different level of detail. Availability of all relevant cost drivers cannot be
assumed. However, the essential parameters must be provided, or estimated, by the customer in
order to perform a supply chain study. In case the customer does not provide the requested data in
terms of quantity or quality, assumptions must be made by the team that carries out the actual cost
estimation. In order to perform good estimations, from the analysis in Chapter 2 can be concluded
that customers must at least be able to provide throughput- and costs-data. Otherwise, there is not
much to fall back on.

16
4 COST ESTIMATION METHODS
After having identified the most important cost drivers of warehouse costs and determined the kind
of data that may be expected from customers, relevant cost estimation methods are described and
evaluated.

In this chapter insights from scientific literature are described and will be evaluated based on the
requirements set in Chapter 1. The goal of this chapter is to identify and describe suitable methods
for cost estimation and thereby answering research question 3:

3. What relevant methods are available to do cost estimations?

The first section gives an overview of the most widely used estimation methods, of which a selection
is discussed in the remainder of this chapter.

4.1 OVERVIEW METHODS

In this section a brief overview of available estimation methods is given, as set out by Leonard (2009).
In his book he describes five and compares three cost estimation methods, of which the comparison
can be found in Table 4.1-1. These main cost estimation methods are:

- Analogy
- Engineering build-up
- Parametric estimating

In addition to these methods, Leonard adds:

- Expert-opinion
- Extrapolation

These five methods will be discussed in the following sections; after which these will be evaluated. All
the methods are discussed in this chapter, with specific methods per subject.
Table 4.1-1: Three cost estimation methods compared (from Leonard, 2009)

17
4.2 PARAMETRIC ESTIMATING
In this section a number of parametric estimation in general is described and several methods are
discussed: regression methods, clustering and machine learning.

According to Leonard (2009), the goal of parametric estimating is to create a statistically valid cost
estimating relationship using historical data. In order to achieve this goal, it is always essential to have
an adequate dataset. To develop a parametric cost model, cost estimators must be determined that
most influence cost. After that, the parametric cost model can be developed with a mathematical
expression, which can range from a simple rule of thumb to a complex regression equation. Dysert
(2008) emphasizes the advantages of parametric estimating as being:

- Efficient: Less time-consuming than more detailed techniques

- Objective: Completely quantity-based
- Consistent: Provides a consistent estimate format and estimate documentation
- Flexible: Wide range of applications and models can be easily adjusted
- Defensible: Able to provide key statistical relationships and metrics for comparison with other
projects

A drawback, as mentioned by Dysert (2008), of parametric estimating is that the quality of the
resulting model cannot be better than the quality of the data it is based upon. In addition, Leonard
(2009) adds that the program attributes being estimated must fall within the range of the dataset.
Otherwise estimation quality is not guaranteed.

The fit of parametric estimation methods with the actual data is in many cases defined using R2, the
‘coefficient of determination’, which is a measure of how well the formed equation explains the
variability of the data (Dysert, 2008). The formulation of R2 looks as follows (Myers, 1986):
𝑛
2
(𝑦̂𝑖 − 𝑦̅𝑖 )2
𝑅 =1− ∑ (4.1)
(𝑦𝑖 − 𝑦̅𝑖 )2
𝑖=1

with 𝑦̅𝑖 being the average of the observations, 𝑦̂𝑖 being the fitted least squares line and 𝑦𝑖 being the
actual observations.

Clearly, R2 is between 0 and 1 and the upper bound is achieved when the model fits the data perfectly.
When multiple explanatory estimators enter the equation, R2 must be corrected for automatic
increasing. This can be achieved by calculating the adjusted R2:
𝑛−1
𝑅̅ 2 = 1 − (1 − 𝑅 2 ) ∗
𝑛−𝑝 (4.2)
with n being the number of observations and p being the number of variables (Dysert, 2008).

4.2.1 Simple linear regression

The first method discussed is simple linear regression. The term ‘regression’ can be traced back to
Galton (1885), who demonstrated that offspring do not tend towards the size of parents but rather
toward the average of both parents. The term’s application in this case was the ‘regression towards
mediocrity’ of offspring. The algebraic procedure of fitting linear equations to data can be traced back
to Carl Friedrich Gauss and Legendre (1805), who first published about the method of least squares.

Simple linear regression is the simplest regression structure; which mathematical formulation is as
follows:

18
 𝑦 = 𝛽0 + 𝛽1 ∗ 𝑥 + 𝜀 (4.3)
where y is the estimator, 𝛽0 and 𝛽1 are the intercept and the slope and 𝜀 is the model error (Myers,
1986). The widely used method to estimate 𝛽0 and 𝛽1 is the least squares method (Legendre, 1805),
that aims to residual sum of squares:

with the fitted value being:

𝑦̂𝑖 = 𝛽0 + 𝛽1 ∗ 𝑥𝑖 (4.4)
and by minimizing

∑𝑛𝑖=1(𝑦𝑖 − 𝑦̂𝑖 )2 (4.5)

The minimization is usually done by assuring the derivative is zero, so by satisfying:

𝑛
𝜕
[∑(𝑦𝑖 − 𝑏0 − 𝑏1 ∗ 𝑥𝑖 )2 ] = 0 (4.6)
𝜕𝑏0
𝑖=1

and
𝑛
𝜕
[∑(𝑦𝑖 − 𝑏0 − 𝑏1 ∗ 𝑥𝑖 )2 ] = 0 (4.7)
𝜕𝑏1
𝑖=1

with 𝑏0 and 𝑏1 being the variables in these equations (Myers, 1986).

Example 1:
The packaging distributor is exploring its data and wants to find out whether the building area of a
warehouse is a good estimator for the total costs. This test is also applied in section 2.2.1. In order to
do this, the packaging distributor applies simple linear regression, as to be seen in Figure 4.2-1.

Figure 4.2-1: Regression line of building area versus the total costs (implemented in AIMMS)

The regression line in this scatterplot represents a straight line through the data with minimal total
distance to the observations. In Figure 4.2-1 the packaging distributor can now easily see how the

19
building area relates to the total costs. When putting the defined values in the regression equation,
the equation and coefficient of determination look as follows:
Table 4.2-1: Regression equation based on building area

Regression line 𝒃𝟎 𝒃𝟏 𝒙𝟏 R2
̂𝒊
𝒚 671,610 11.12 Building area 0,85

The fit between the building area and the total costs seems to be strong, since the R2 is close to 1.
From this can be concluded that building area might be a good estimator for the total costs.

4.2.2 Multiple linear regression

Multiple linear regression is an extension of simple linear regression. Equation 4.3 can easily be
expanded to a multiple linear regression model.

𝑦 = 𝛽0 + 𝛽1 ∗ 𝑥1𝑖 + 𝛽2 ∗ 𝑥2𝑖 + ⋯ + 𝛽𝑘 ∗ 𝑥𝑘𝑖 + 𝜀𝑖 (4.8)

with 𝑘 explanatory estimators.

According to Myers (1986) the main differences with simple linear regression are that the data is no
longer easily displayed (it may have more than two dimensions) and the least squares method is more
complicated and requires more development. For larger models, with multiple regressors, the least
squares method is often applied using statistical software as R or SAS (Montgomery & Peck, 2015).
Even in cases where the number of estimators is as small as 3, an explicit solution could get very
complicated (Fox, 2008).

Montgomery & Peck (2015) further mention that adding an important predictor to a regression model
can often result in a much better fitting model with a smaller standard error. As a consequence,
narrow confidence intervals and narrower prediction intervals. In addition, multiple linear regression
models are often used as empirical models or approximating functions.

Mostly due to the relatively simple implementation and intuitive results of linear regression,
applications of linear regression occur in almost every field, including engineering, physical and
chemical sciences, economics, etcetera. Regression analysis is probably the most widely used
statistical technique (Montgomery & Peck, 2015).

Example 2:
Suppose that the packaging distributor is not satisfied with the cost estimation, only based on
building area (as in Example 1). Therefore, an additional explanatory estimator is added to the
equation: throughput. The multi-regression equation and the (adjusted) coefficient of determination
look as follows:
Table 4.2-2: Regression equation based on throughput and building area

Regression line 𝒃𝟎 𝒃𝟏 𝒃𝟐 𝒙𝟏 𝒙𝟐 R2 Adj. R2

̂𝒊
𝒚 186.46 3.16 0.04 Building Throughput 0,923 0,921
area

When looking at the adjusted R2 and compare it to the simple linear regression with only building area
as estimator, the fit is improved. This means that the multi-regression equation is likely to give better
estimations.

20
4.2.3 Nonlinear regression
In the earlier described regression models, the structure is linear in the model coefficients. In many
scientific fields, knowledge about the experimental situations suggests the use of a less empirical,
more theoretically based, nonlinear model. Nonlinear regression models can have many forms, but all
nonlinear models have one thing in common: at least one of the parameters enters the model in a
nonlinear way (Myers, 1986). In fact, a nonlinear equation can be of the following form:

𝑦 = 𝛽0 + 𝛽1 ∗ 𝑥 𝑣 + 𝜀 (4.9)

This is the same equation as for simple linear regression, but it includes an exponent v. It totally
depends on the situation what value v is.

In general, if the data does not produce a linear fit, nonlinear regression can be used in order to see if
the relations are nonlinear (Leonard, 2009). Typically, a nonlinear regression model adds complexity
in comparison to linear models. Complexity may help in cases the forecaster has profound knowledge
of the situation and improves the ability to fit historical data, but it often harms forecast accuracy
(Armstrong, 2001).

Example 3:
The packaging distributor thinks it is likely that throughput is the best estimator and wants to use it
as only variable in the regression equation. The packaging distributor believes that the relation
between throughput and costs is not strictly linear, because economies of scale occur. Therefore,
the regression line is typically a deflecting line with an exponent smaller than 1. When fitting the
data using nonlinear regression of the form of Equation 4.9, it looks as follows:
Table 4.2-3: Regression equation based on throughput

Regression line 𝒃𝟎 𝒃𝟏 𝒙𝟏 𝒗 R2
̂𝒊
𝒚 -37,417 2.13 Throughput 0.79 0,91

The assumption of the packaging distributor was right and it seems that a nonlinear regression
equation better fits the data and is likely to provide better estimations.

Figure 4.2-2: Nonlinear regression line of throughput versus the total costs (implemented in AIMMS)

21
4.2.4 Clustering
The cluster analysis problem is the last step in the progression of category sorting problems. In
classification the category structure is known and the objective is to discover a structure that fits the
observations (Anderberg, 2014). In many cases, subgroups within data exist that have different
regression functions. It can be extremely difficult or even impossible to sort out these observations
into groups and clusters, especially in situations involving several independent variables (Desarbo,
1989). Backer & Jain (1981) define clustering as follows: “In cluster analysis a group of objects is split
up into a number of more or less homogenous subgroups on the basis of an often subjectively chosen
measure of similarity, such that the similarity between objects within a subgroup is larger than the
similarity between objects belonging to different subgroups.”

Clustering is a form of unsupervised classification, which means that observations are not labeled yet;
so the goal of clustering is to separate an unlabeled data set into a finite and discrete set of ‘natural’,
hidden data structures (Xu, 2005).

Figure 4.2-3: Clustering procedure (from Xu, 2005)

A typical cluster procedure is described by Xu (2005) and is illustrated in Figure 4.2-3. Feature selection
chooses distinguishing factors from a set of candidates, while feature extraction uses transformations
to generate new features from the original ones. The clustering algorithm design is concerned with
the selection of a proximity measurement and the construction of a criterion function. In the cluster
validation step the formed clusters are analyzed, as well as the cluster algorithm. In the final stage,
the interpretation of the formed clusters takes place, to see whether the users are provided with
relevant insights from the initial data.

Over time different clustering algorithms have been developed. The simplest and most popular
algorithms are single (Sneath, 1957) and complete linkage techniques (Sorensen, 1948), both
examples of hierarchical clustering methods. Single linkage techniques, also called nearest neighbor
method, determines the distance between clusters as the distance of the closest observations
between two clusters. Complete linkage techniques determine the distance between clusters as being
the distance of the farthest observations between two clusters. A major disadvantage of these
methods is that they lack robustness and are sensitive to noise and outliers. Furthermore, these
algorithms do not correct for previous misclassification (Xu, 2015).

Squared error-based clustering, based on assigning observations into K clusters can be achieved using
total enumeration, but other, more efficient methods are available (Xu, 2015). K-means clustering is
the most widely known squared error-based cluster algorithm, which is an iterative algorithm that
repeats the reassignment of observations to the nearest cluster until no change of observations within
each cluster takes place (MacQueen, 1967). Advantages of the K-means are that the method is very
simple and can easily be implemented in solving many practical problems. Disadvantages are that
there is no efficient or universal initialization method, convergence is not guaranteed and it is sensitive

22
to outliers and noise. Therefore, many variants of K-means have been developed that overcome these
disadvantages and improve performance (Xu, 2015).

Many other clustering techniques are developed over time, but the general idea of clustering is made
clear at this point already. The best example of the application of clustering can be found in Section
2.4.3.

4.2.5 Machine learning

A more advanced form of parametric estimating is machine learning. Breiman (2001) compares
traditional data analysis methods as (multi-) regression with algorithmic modelling (machine learning
techniques). According to Breiman, the concepts of both methods can be best illustrated as in Figure
4.2-4. The main difference between both concepts is that the traditional data analysis forms a model
estimated from the data, while algorithmic modeling considers the inside of the ‘box’ complex and
unknown.

Figure 4.2-4: Traditional data analysis versus algorithmic modeling (From Breiman, 2001)

Michalski et al. (2013) distinguish 5 different forms of machine learning:

- Rote Learning: Machine stores new knowledge in memory.

- Learning from instruction: Acquiring knowledge from a teacher or other organized source,
requiring the learner to transform the knowledge from input language to internally-usable
format.
- Learning by analogy: Acquiring new facts or skills by transforming and augmenting current
knowledge with strong similarity to the desired concept into a form effectively useful in the
new situation.
- Learning from examples: Given a set of examples of a concept, the learner acquires a general
concept description that describes the relations within the data. Learning from example is a
form of inductive learning.
- Unsupervised learning: Acquiring underlying relationships from unlabeled data.

As learning from examples is most applicable to this research, the focus will be on these forms, of
which a few examples of machine learning techniques will be discussed.

- Bayesian Networks are directed probabilistic graph models, that allow efficient and effective
representation of the joint probability distribution over a set of random variables. The
conditional probabilities are learned from training data. Surprisingly simple Bayesian
classifiers, with strong assumptions of independence among features, called naïve Bayes, is
competitive with state-of-the-art classifiers. (Friedman et al. 1997).
- Regression trees, a specific form of decision trees, automatically decide on the splitting
variables and split-points, and what shape the tree should have. The size of the three is
important, since large trees can imply overfitting, where small trees might not capture the
important structure (Friedman et al. 2008).

People make mistakes during analysis, trying to establish relationships between multiple features and
find it difficult to find solutions to certain problems. Machine learning techniques can be successfully

23
applied to these problems, improving efficiency (Kotsiantis, 2007). As Breiman (2001) states,
algorithmic models like random forests (extension of a tree) and Bayesian methods provide generally
more accurate results than traditional data analysis methods.

Example 3:
The packaging distributor has big loads of data; a large set of observations with many possible
explanatory variables. In order to make good cost estimations the packaging distributor thinks
machine learning might be a good option. ORTEC develops an inductive learning algorithm that
defines inter-relations between the variables using a decision tree-structure. An example of how the
first ‘leafs’ of such an algorithm may look like can be found in Figure 4.2-5.

Figure 4.2-5:Example of how a decision tree for cost estimation may look like

4.2.6 Activity based costing

Activity based costing (ABC) systems are used to allocate indirect costs to processes in a way that
accurately reflects how costs are actually incurred (Richards, 2010). According to Cooper & Kaplan
(1988), an activity based costing system can represent product costs relatively different from data
generated by traditional systems. Cooper & Kaplan further mention that all of a company’s activities
exist to support the production and delivery of goods and services, which therefore must all be
considered as product costs.

One of the main reasons of applying ABC is that it allows management to establish priorities, by
focusing on improving the value-added activities where the largest savings can be realized. (Angelis
& Lee, 1996).

An extension of traditional ABC, is time-driven ABC, where companies use time equations to estimate
the time spent on each activity (Everteart et al., 2008). The same authors mention that traditional ABC
oversimplifies a large part of the activities and misallocates a high percentage of the indirect costs.
The results of their study shows significantly better results using time-driven ABC relative to traditional
ABC. Taken this into account, a major disadvantage of time-driven ABC is that the measurement of
durations is very laborious or sometimes even impossible (Varila et al., 2007).

ABC is, in essence, an accounting method, but once all costs are assigned to the products (or another
entity); it may be a very useful and a relatively simple tool to estimate costs of new locations.

Example 4:
The packaging distributor already applies activity based costing. It allocates all the costs associated
with the warehouses to the throughput, resulting in costs per product; in this case 5 cents per
product. To provide an estimation, the throughput of a warehouse only has to be multiplied with the
costs per product.

24
4.3 ENGINEERING BUILD-UP
The engineering build-up method develops the cost estimate at the lowest level of a warehouse work-
breakdown-structure. It takes into account one piece at a time and eventually, the sum of the pieces
becomes the estimate. The engineering build-up is also referred to as ‘bottom-up estimate’. The
engineering build-up estimate is done at the lowest detail level and mainly consists of labor and
material costs, further increased by overhead (Leonard, 2009).

The engineering build-up method can also be used to enhance historical cost data by adding data
points that are needed for further analysis. The engineering build-up model must be formulated from
a combination of engineering experience and judgment and experience within the particular domain
of the system in question. It seems essential that when setting up this cost model, much care and skill
are employed (Hart et al., 2012).

4.3.1 CCET
Within ORTEC, the engineering build-up method is thoroughly used within the CCET-department. CCET
stands for Capital Cost Estimating Tool, which is typically used to make accurate estimations of new
refineries. For this purpose, loads of historical data are available and used for analysis. Estimations
can be made on different levels, from a rough cost estimation to a very detailed and specific cost
estimation.

4.4 ANALOGY
The use of analogy as cost estimation method is mainly used when there are little observations
available. Analogy typically relies for a large portion on expert opinion to modify the existing data to
make an approximation. Analogies are often used as a validation method. When a data analysist is
using a more detailed cost estimating method, an analogy can provide a useful logical check (Leonard,
2009). As Hart et al. (2011) mention, the analogous costing method estimates the cost of a product or
system, based on the differences from similar products or systems.
Table 4.4-1: Example of the analogy cost estimating method (from Leonard, 2009)

When the analogy is strong, the estimate is likely to be strong as well. A major disadvantage is that
the analogy relies heavily on expert opinion in case the observation used for the analogy differs in
some aspects from the new system (Leonard, 2009).

4.5 EXPERT OPINION

In some cases, little or no data is available, which means quantitative analysis is not possible or
sufficient. In these cases, expert opinions can be solely used to do estimations or be added to set up
a cost model. The latter case is often applied. As earlier mentioned in this chapter, expert opinion is
sometimes required to build up a cost model; it can be of use to initialize a model by, for example,
estimating the probabilities in a Bayesian network, identifying the cost drivers or setting up an
analogy. Expert opinion can be subdivided into different methods, of which structured judgment is
superior to other ‘unstructured’ methods. Structuring the interviews with care enhances reliability
and validity (Armstrong, 2001).

25
Especially in the field of sales forecasting, expert opinion based estimation is the most commonly
applied method. Despite its popularity, especially within sales-forecasting, it may not always be the
best method. More quantitative methods may be preferred when available and at worst, quantitative
methods seem to be as accurate as judgmental methods (Armstrong, 2001).

4.6 SUMMARY COST ESTIMATION METHODS

In the sections above, several cost estimation methods have been described. Parametric cost
estimating methods discussed consist of different kinds of regression analysis, clustering and machine
learning. Regression methods do not need very much observations or variables, in contrast to machine
learning which ideally needs many observations and many variables to have an advantage over the
less advanced regression methods. Parametric estimation methods are very suitable for cost
estimation and relatively fast to apply. The power of the estimation is totally dependent on the quality
of the data and the inner relations.

Activity based costing is the simplest and easiest applicable cost estimation method discussed. The
advantage of this method is that it is very intuitive and the needed KPI’s are many times already
available since most companies already apply some kind of cost allocation.

The engineering build-up has the advantage of providing very accurate cost estimations and the
disadvantage of being very time-consuming.

The analogy provides good cost estimations when the analogies are strong. It is, in many cases, hard
to find a similar case and if this is not possible the analogy relies heavily on expert opinion.

Expert opinion is especially useful as additional technique to initialize a model or fill in assumptions
when data is missing. When no data is available, it is possibly the most suitable cost estimation method
but when useful data is available, expert opinion is not preferred over quantitative methods.

4.7 METHOD COMPARISON AND CHOICE

After having identified different cost estimation methods and having defined their strengths and
weaknesses, this section provides in evaluating and selecting methods to apply to a case study. The
described methods are compared to each other and reviewed taking into account the requirements
and limitations as stated in chapter 1. The fourth research question will be answered in this section:

4. Which method or methods are most suitable given the situation at ORTEC?

With the problem statement, as given in chapter 1, several requirements are set. In order to make a
good comparison between the described methods, the amount (number of observations) and quality
(number of variables) are added; criteria that have to do with data availability. The implementation
complexity is taken into account as well. Thus, the estimations are compared based on the following
criterions:

- Observations needed: The less observations are needed for a good estimation, the better.
- Variables needed: The less variables needed, the better.
- Accuracy: the goal is set at 10 percent deviation in 90 percent of the cases.
- Calculation time: the goal is set to make an estimation in the order of seconds
- Generosity: The model must be applicable to every sector
- Complexity: The lower the complexity, the easier to develop and the easier to explain to
stakeholders.

26
The results of this comparison are summarized in Table 4.7-1. The estimation methods are analyzed
per criterion:

Observations needed: The parametric methods and the engineering build-up all need a substantial
amount of observations in order to do accurate estimations (>10 observations), with machine learning
needing much more (>100 observations) in order to have additional value. The analogy typically needs
only one observation and the expert opinion based method none.
Table 4.7-1: Different estimation methods with their score per criterion
Observations needed

Variables needed

Accuracy

Calculation time

Genericity

Complexity
Criterion →

Method ↓
Machine learning +++ +++ +++ ++ +++ +++
Linear regression ++ + ++ + +++ +
Nonlinear regression ++ + ++ + +++ +
Clustering ++ + ++ + +++ +
ABC ++ + + + +++ +
Engineering Build-up ++ ++++ +++ +++ + ++
Analogy + ++ + ++ + +
Expert opinion - - + +++ + ++

Variables needed: In order to be effective, machine learning methods need many variables, otherwise
normal regression methods are already sufficient. The engineering build-up needs an endless number
of variables, as every detail of het new warehouse needs to be modelled. All other methods can
provide estimations with only one variable, but the estimation power may grow with additional
variables.

Accuracy: Machine learning and the engineering build-up are expected to be the most accurate
estimation methods, followed by the regression methods and clustering. ABC, expert opinion and
analogies, are expected to provide less accurate estimations.

Calculation time: The engineering build-up is typically the most time-consuming process, followed by
expert opinion. Machine learning can take time to set-up, sometimes in combination with expert
opinion, and the running time increases with the number of variables included. Analogies typically
need some expert opinion, which makes this process take more time. The other methods can provide
an estimate within the order of seconds.

Genericity: Once built, machine learning, regression, clustering and ABC can be applied to every
dataset. The engineering build-up, the analogy and the expert opinion method are custom, so they
need to be set up for every individual estimation.

Complexity: Machine learning methods can be difficult to build and sometimes it is even harder to
interpret the results. The engineering build-up is complex in the sense that everything needs to be
modelled in a detailed way, while for the expert opinion method the complexity lies in structuring the

27
interviews and implementing the results. The remaining methods are all easy interpretable and
relatively easy to set up.

In Figure 4.7-1 the different methods are set out in a matrix, with the number of variables and
observations on the y-axis and accuracy on the x-axis.

Figure 4.7-1: Performance matrix of different estimation methods

When looking at this matrix, the most advantageous methods can obviously be found in the bottom
right corner. Since the data quality and quantity are both not expected to be high, the methods in the
upper right corner are out of scope of this research. Expert opinion based techniques and analogies
are too time-consuming and are out of scope as well. This means that the remaining methods are
being tested within a case study in order to evaluate their estimating performance for the given
dataset.

4.8 CONCLUSION
This Chapter provides in answering two research questions:

3. What relevant methods are available to do cost estimations?

4. Which method or methods are most suitable given the situation at ORTEC?

After having identified, described and analyzed different cost estimation methods, four methods
match the requirements and fall within the scope of this research. These methods are:

- ABC
- Linear regression
o Simple linear regression
o Multiple linear regression
- Nonlinear regression
- Clustering

The implementation and mathematical formulation of these methods, as well as adjustments and
additions, are described in the following chapter. Since all methods are implemented in AIMMS and
some are solved using the optimization engine, the problem classification is added as well.

The findings of this chapter will be further evaluated in a case study, performed in Chapter 6, to
determine which method provides the most accurate cost estimations.

28
5 APPROACH & MATHEMATICAL FORMULATION
The chosen methods, as described in the previous chapter, are all applied to the dataset of the
packaging distributor, to determine the estimation power and decide what cost estimation method is
the most suitable. In order to put these methods to use, they have to be formulated and implemented.
One of the requirements set by ORTEC is that the cost analysis and estimation find place within
AIMMS. Since AIMMS is especially good at solving optimization problems, the methods are all, if
possible or needed, formulated as an optimization problem.

Furthermore, in this chapter, the cost estimation methods are formulated mathematically and, if
needed, explained in more detail. First, the formulation of the linear regression methods is described,
followed by an extension of this, to take into account additional factors. In this extension, called
‘extended regression’, the country factor, the automation level and the area factor are all
implemented in the regression model. Further, the clustering method, as well as the nonlinear
regression and ABC are formulated.

All methods are formulated in the same format:

- Sets and indices: a dimension of a parameter or variable (e.g. observations)

- Parameters: input data (e.g. the throughput per observation
- Variables: the value to be determined (e.g. the slope in a regression equation)
- Constraints: equations with both variables and parameters that bound variables (e.g. the
slope must be greater than 0)
- Problem statement: objective function with both variables and parameters, this equation is
to be optimized (e.g. minimize total costs)

5.1 SIMPLE LINEAR REGRESSION & MULTI-REGRESSION

The simple linear regression and the multi-regression are implemented in AIMMS, in which the
regression line is determined by minimizing the squared error; thus by using the optimization engine
of AIMMS. The formulation of both simple linear regression and multi-regression is the same; if the
number of estimators in the model is set to 1, the equation equals the simple linear regression
equation.

The mathematical formulation is then as follows:

5.1.1 Sets and indices

 Set I of all observations i;
 Set К of all cost drivers k;

5.1.2 Parameters
 Estimator xi,k for all cost drivers κ and for all observations i;
 Total costs yi for all observations i;

5.1.3 Variables
- Slope S k for all cost drivers k;
- Intercept β;

5.1.4 Problem statement

2 (5.1)
min ∑𝑖( (𝑥𝑖,𝑘 ∗ 𝑆𝑘 + 𝛽) − 𝑦𝑖 )

29
5.1.5 Problem classification
The mathematical problem as stated is a quadratic problem, which can be solved with CPLEX, an
integrated linear solver integrated in AIMMS which provides instant optimal solutions, if feasible.

5.2 EXTENDED REGRESSION

The cost driver analysis in Chapter 2 resulted in several cost drivers, of which the country and the
location are hard to take into account in a regression equation, because they are not continuous or
ordinal. Normalizing the data based on wage differences alone did not provide a better fit, but to
account for the country and location factor, a conceptual regression equation is set up, based on three
elements that are all driven by throughput: labour costs, building area costs and throughput costs.

Labour costs: The amount of labour needed is a combination of the automation level and the
throughput. The subsequent costs of the amount of labour then is mainly dependent on the country
the warehouse is in. In order to take this into account, the regression slope is corrected per
automation level:

Figure 5.2-1: Cost drivers of labour costs

Building area costs: The cost of square meters is country- and location-specific, with location being
urban or rural. The needed building area of a warehouse is assumed to be mainly driven by the
throughput, which means that the costs can be modelled with a regression slope corrected per
location-type.

Figure 5.2-2: Cost drivers of building area costs

30
The idea behind this regression equation is that the country factor, the automation factor and the
location factor are determined by the data provided. All factors are variables that can take values
between 0 and 1, that represent their relative influence.

The mathematical formulation is then as follows:

5.2.1 Sets and indices

 Set I of all observations i;
 Set К of all cost drivers k;
 Set F of all countries f;
 Set A of all automation levels a;
 Set H of all locations h;

5.2.2 Parameters
 Estimator 𝑥𝑖𝑇𝑃 of the throughput per observation i;
 Total costs yi for all observations i;

5.2.3 Variables
- Capacity slope 𝑆 𝑇𝑃 ;
- Automation slope 𝑆 𝐴𝑢𝑡𝑜 ;
- Location slope 𝑆 𝐿𝑜𝑐 ;
- Intercept β;
- Country factor, dependent on country f and observation i: 𝜆𝐹𝑖 ∈ [0, 1] ;
- Location factor, dependent on location h and observation i: 𝐺𝐻𝑎𝑟𝑒𝑎
𝑖
∈ [0, 1] ;
- Automation factor, dependent on automation level a and observation i: 𝑀𝐴𝐴𝑢𝑡𝑜
𝑖
∈ [0, 1] ;

5.2.4 Problem statement

The standard regression equation is extended, with different factors. The simple regression equation
of Section 5.1 driven by one cost driver (in this case throughput) is extended with a country factor.
Labour costs are introduced, driven by throughput corrected by both the country and automation
level. The building area costs are taken into account, driven by throughput as well and corrected by
both the country and the area an observation is in. Throughput costs, labour costs and building area
costs all have their own slope which will be equal to 1 in case the factor does not have any impact.
The same holds for the correction factor; if for example the automation factor is not of any influence,
the value of this variable will be 1 for all automation levels.

These three equations are all implemented in one single objective. Since all three parts are driven by
throughput and corrected by the country factor, these are written on the outside of the equation in
the objective.

- Throughput costs:

𝑥𝑖𝑇𝑃 ∗ 𝜆𝐹𝑖 (𝑆 𝑇𝑃 ) (5.1)

- Labour costs:

𝑥𝑖𝑇𝑃 ∗ 𝜆𝐹𝑖 (𝑆 𝐴𝑢𝑡𝑜 ∗ 𝑀𝐴𝐴𝑢𝑡𝑜

𝑖
) (5.2)
- Building area costs:

𝑥𝑖𝑇𝑃 ∗ 𝜆𝐹𝑖 (𝑆 𝐴𝑟𝑒𝑎 ∗ 𝐺𝐻𝑎𝑟𝑒𝑎

𝑖
) (5.3)

31
Objective:
2 (5.4)
𝑇𝑃
𝑆 +
min ∑𝑖 𝑇𝑃
𝑥𝑖 ∗ 𝜆𝐹𝑖 (𝑆 𝐴𝑢𝑡𝑜
∗ 𝑀𝐴𝐴𝑢𝑡𝑜
𝑖
+) + 𝛽 − 𝑦𝑖
𝐴𝑟𝑒𝑎 𝑎𝑟𝑒𝑎
𝑆 ∗ 𝐺𝐻𝑖
( )
( )
5.2.5 Problem classification
The multiplication of different variables with each other and the quadratic term, make this problem
nonlinear, which is solved using IPOPT, providing acceptable local optima.

5.3 NONLINEAR REGRESSION

Nonlinear methods can have many forms, of which the simplest application is adding an exponential
component to the estimator. Instead of using a fixed exponent, it is in this implementation applied as
variable and is defined using the optimization engine of AIMMS.

The mathematical formulation of this application is as follows:

5.3.1 Sets and indices

 Set I of all observations i;
 Set К of all cost drivers k;

5.3.2 Parameters
 Estimator xi,k for all cost drivers κ and for all observations i;
 Total costs yi for all observations i;

5.3.3 Variables
- Slope S k for all cost drivers k;
- Intercept β;
- Power variable V;

5.3.4 Problem statement

2 (5.5)
min ∑𝑖( (𝑥𝑖,𝑘 𝑉 ∗ 𝑆𝑘 + 𝛽) − 𝑦𝑖 )

5.3.5 Problem classification

The introduction of an exponential component makes this problem shift from a relatively simple
quadratic program to a nonlinear program, which is harder to solve. The used solver for this problem
is IPOPT, which does not provide a global optimum. It does provide an acceptable solution after 3000
iterations.

5.4 CLUSTERING
Cluster methods are mostly implemented as an iterative procedure, with observations exchanged
between clusters, but optimization is a suitable method as well and presumably faster. The problem
definition does not differ much from that of multi-regression, except for the introduction of an extra
index, with as cardinality the number of clusters as defined beforehand, and a binary variable that
states whether an observation is within a cluster. The objective is, except for the binary variable, the
same as for multi-regression: minimizing the sum of the squared error.

32
The number of clusters used is determined using an iterative procedure. This procedure first calculates
the maximum amount of clusters possible, given that each cluster must have at least 10 observations.
After that, the procedure iterates with a shifting number of clusters. The number of clusters for which
the sum of the squared error is minimal is used. Therefore, it may occur that the estimations are the
same as for linear regression: when the ideal number of clusters is 1.

The estimation input is then compared to the average within each cluster and assigned to the cluster
for which the difference is minimal.

5.4.1 Sets and indices

 Set I of all observations i;
 Set К of all cost drivers k;
 Set C of all clusters c

5.4.2 Parameters
 Estimator xi,k for all cost drivers κ and for all observations i;
 Total costs yi for all observations i;
 Number of clusters;

5.4.3 Variables
- Slope S k,c for all cost drivers κ and all clusters c;
- Intercept βc for all clusters c;
- Binary variable that states whether observations i is in cluster c: 𝛿𝑖,𝑐 ∈ {0,1};

5.4.4 Problem statement

2 (5.6)
min ∑𝑐 ∑𝑖 𝛿𝑖,𝑐 ∗ (∑𝑘(𝑥𝑖,𝑘 ∗ 𝑆𝑘,𝑐 ) + 𝛽𝑐 ) − 𝑦𝑖 )

s.t. ∑ 𝛿𝑖,𝑐 > 10 (5.7)

𝑖
∑ 𝛿𝑖,𝑐 = 1 (5.8)
𝑐

5.4.5 Problem classification

The introduction of a binary variable, makes the problem shift from a quadratic problem to a mixed
integer nonlinear problem (MINLP). This kind of problems are typically difficult to solve and feasible
solutions are found by relaxing the binary variable: 𝛿𝑖,𝑐 ∈ [0,1] and round the solution to the nearest
integer. The subsequent NLP is solved using the IPOPT-solver, which provides acceptable local optima.

5.5 ABC
Activity based costing is, in essence, just assigning costs to an activity. The activities available for the
cost estimation of warehouses are the quantitative cost drivers, such as throughput or building area.
The mathematical formulation, is therefore quite simple:

5.5.1 Sets and indices

 Set I of all observations i;
 Set К of all cost drivers k;
 Set F of all countries f;

33
5.5.2 Parameters
 Estimator xi,k for all cost drivers κ and for all observations I;
 Total costs yi for all observations i;
- Slope S k for all cost drivers k;

5.5.3 Problem statement

𝒚𝒊 (5.9)
𝑺𝒌 = ∑
𝒙𝒊,𝒌
𝒊

When, for example, distinguishing between the country an observation is in, adding an index is
sufficient. The same holds for other indices, such as automation level or location.
𝒚𝒊,𝒇 (5.10)
𝑺𝒌,𝒇 = ∑
𝒙𝒊,𝒌,𝒇
𝒊

5.5.4 Problem classification

This problem is just a straight calculation, so no optimization or iteration is involved. AIMMS
automatically calculates the values through definitions.

5.6 CONCLUSION
In this chapter, the estimation methods used for the case study are worked out mathematically in the
way they are implemented in AIMMS. In addition to the selected methods, an extra estimation
method is introduced. This method is an extension to simple linear regression and takes into account
the influence of the country, the area and the automation level with a correcting factor. The idea
behind this is that, based on these cost drivers, the building area and labour costs can be accounted
for in addition to the throughput costs

All methods, except ABC, are implemented using optimization. Some methods are not easily solved,
but to give fast estimations, which is one of the requirements, these are solved to local optima within
a few seconds. The local optima provided are proven to be acceptable and near-optimal.

In the next chapter a case study is performed, where all the formulated methods are applied to the
data of the packaging distributor, to find out which method is the most suitable for cost estimation.

34
6 CASE STUDY
To put the insights gained in this study to the test, a case study is conducted in this chapter, thereby
continuing in answering the following research question:

4. Which method or methods are most suitable given the situation at ORTEC?

The dataset of the earlier introduced packaging distributor is used to test the different cost estimation
methods as described in Chapters 4 and 5. First, the set-up of the case study is discussed, followed by
the testing procedure. The results of the performed tests are presented and the most accurate
methods are discussed in more detail. At the end of this chapter, conclusions are drawn about the
performance of the different cost estimation methods.

6.1 SET-UP CASE STUDY

To check the performance for all the methods the dataset must contain enough data to perform
analysis. The dataset used for this analysis is that of the packaging distributor, of which the
characteristics are discussed in Chapter 1. This dataset contains 87 observations of warehouses all
over Europe, based in 16 different countries. As earlier mentioned in chapter 2 and recaptured in
Table 6.1-1, this dataset does not contain all the cost drivers for an ideal test-scenario but is sufficient
to test the different proposed estimation methods. The main side note is that the extended regression
method, as described in the previous chapter can only be applied partly.
Table 6.1-1: Independent variables present within the dataset of Customer 2

Observations Building #Employees Automation Throughput Country Location Total

area level costs
87 x x x x

6.2 TESTING PROCEDURE

All the estimation methods are being tested on the basis of throughput. This has three main reasons:

- The first reason is that the cost driver analysis resulted in throughput as being the most
important cost driver.
- The second reason is that throughput is typical input for a supply chain study, while building
area is not.
- The third reason is that the most observations contain throughput. This is not the case for the
building area.

The only exception is the multi-regression, in which the building area is taken into account in the
regression equation.

An overview of the methods, the used cost driver (predictor) and additional used data is shown in
Table 6.2-1.
Table 6.2-1: Estimation methods and used predictors

Method Cost driver Additional data

Linear regression Throughput
ABC Throughput
ABC per Country Throughput Country
Extended regression Throughput Country

35
 Normalized Linear regression Throughput Wage-index
Nonlinear regression Throughput
Multi-regression Throughput + Building
area

Due to the limited amount of observations within the dataset that contain the automation level and
no observation containing information about the area, only a part of the extended regression formula
is used. The objective can therefore be reduced to:
2
min ∑𝑖(𝑥𝑖𝑇𝑃 ∗ 𝜆𝐹𝑖 ∗ 𝑆 𝑇𝑃 + 𝛽 − 𝑦𝑖 ) (6.1)

To be able to test the extended regression method, the observations per country must be as large as
possible, in order to see the effect of the adjusting country-factor. That is why is chosen for a test set
with three different countries, with a total of 29 observations (see Table 6.2-2).
Table 6.2-2: Number of observations per country used for the test

Country Observations
Germany 7
Spain 13
Netherlands 9
Total 29

Because the dataset is relatively small, the chosen method to define the accuracy of the estimations
is cross-validation (Breiman, 2001). From the dataset, five different test sets are selected, with each
observation only present in one test set. For every test, the total costs of the observations in the test
set are estimated based on the cost model following from the observations in the base set. In the
following section, the results of this testing procedure are presented.

6.3 RESULTS
The absolute results, expressed in the root mean squared error (RMSE), of the cross validation can be
found in Table 6.3-1. The RMSE is defined by first defining the average sum of squares over all the
observations in the test sets and extracting the root of the result. RMSE is also referred to as the
sample standard deviation.
Table 6.3-1: Results of the cross-validation, expressed in RMSE

Method RMSE Rank

Linear regression 1,018,164 5
Clustering 1,100,990 7
ABC 1,171,860 6
ABC per Country 999,677 4
Extended regression 905,488 2
Normalized Linear regression 1,308,426 8
Nonlinear regression 929,619 3
Multi-regression 788,219 1

Obviously, the method that results in the lowest RMSE produces the most accurate estimations.
From Table 6.3 is easily seen that the multi-regression method has the lowest RMSE and therefore

36
provides the most accurate estimations. The sample standard deviation, or RMSE, of respectively
extended regression and nonlinear regression follows on a second and third place. The equations of
these three cost estimation methods can be found in Table 6.3-2. Because the values of the slope,
the intercept and the exponent (nonlinear regression) were relatively stable over the five test sets,
the average of these values is presented. The extended regression equation is using in this
implementation nothing more than just a simple linear regression equation per country but
performs significantly better than simple linear regression over all countries.
Table 6.3-2: Average equation of the three most accurate cost estimation methods

Cost estimation method Equation

Multi-regression 0.4 ∗ 𝑇ℎ𝑟𝑜𝑢𝑔ℎ𝑝𝑢𝑡 + 3,5 ∗ 𝑆𝑢𝑟𝑓𝑎𝑐𝑒 + 102,881
Extended regression Germany 0.037 ∗ 𝑇ℎ𝑟𝑜𝑢𝑔ℎ𝑝𝑢𝑡 + 552,460
Spain 0.034 ∗ 𝑇ℎ𝑟𝑜𝑢𝑔ℎ𝑝𝑢𝑡 + 552,460
Netherlands 0.052 ∗ 𝑇ℎ𝑟𝑜𝑢𝑔ℎ𝑝𝑢𝑡 + 552,460
Nonlinear regression 28 ∗ (𝑇ℎ𝑟𝑜𝑢𝑔ℎ𝑝𝑢𝑡)0.68 − 8,078

When taking a closer look to multi- and extended regression, the slope of throughput seems quite
stable. An approximate cost per unit of 4 cents seems like a good estimation, since the simple linear
regression over all countries also indicated the same slope. So, next to developing a cost estimation
equation this kind of analysis could be very useful for obtaining insight in product costs or fixed costs
(indicated by the intercept in these equations).

6.4 RESULTS WITH RESPECT TO GOAL ORTEC

Now that the different cost estimation methods have been evaluated, the results can be held against
the goal of ORTEC. To recall, the goal of ORTEC is:

- Perform cost estimations with a maximum deviation of 10% in 90% of the cases

When looking at the test results in Figure 6.4-1, it is clear that this objective will not be met. When
looking for a percentage above 90%, nonlinear and multi-regression only succeed when a deviation of
40% is allowed.

37
Figure 6.4-1: Accuracy per method

It can be concluded that cost estimations, solely based on throughput, will not provide sufficiently
accurate estimations. Adding building area (multi-regression) or the country a warehouse is in
(extended regression) to the equation, does not provide more accurate estimations.

6.5 CONCLUSION
The different cost estimation methods have been compared to each other, based on accuracy to give
a final answer to research question 4:

4. Which method or methods are most suitable given the situation at ORTEC?

The three most accurate techniques are multi-regression, extended regression and nonlinear
regression. The most accurate method is nonlinear regression, based solely on throughput.

Although these methods provide the most accurate estimations for this set-up, when looking at the
goal of this research they do not meet the requirement set by ORTEC:

- Perform cost estimations with a maximum deviation of 10% in 90% of the cases

The nonlinear regression equation estimates 39% of the time within 10% deviation and the required
90% is only achieved when increasing the allowed deviation to 40%.

This means that using throughput as main cost driver is not sufficient for performing warehouse cost
estimations, since it does not result in meeting the accuracy requirement as set by ORTEC. Additional
cost drivers as building area and the country a warehouse is in, do not provide higher accuracy. When
the bar is set lower, these techniques may be applicable to do a rough cost estimation and next to
that they may provide insight in differences in product costs per country and typical fixed costs.

In the next chapter, the implementation of the different cost estimation methods is described and in
the last chapter the conclusions and recommendations following from this research are discussed.

38
7 IMPLEMENTATION
The different cost estimation methods are all set out, presented mathematically and eventually tested
for accuracy. In this chapter the implementation of the different methods is discussed, thereby
answering the last research question:

5. How can the insights gained in this research be implemented?

Therefore, in this chapter the implementation, the user interface and the performance of the cost-
estimating methods is discussed.

7.1 IMPLEMENTATION
One of the requirements set by ORTEC is that the warehouse cost estimation is fully implemented in
AIMMS. Therefore, the methods are formulated in such a way they can be effectively applied in
AIMMS, as shown in Chapter 5. For the different regression methods, the optimization engine
integrated within AIMMS is used to define the slope, intercept and/or exponent of a regression line.

The data provided by the customer must be delivered in a specified matrix in Excel, in order to
efficiently load it into AIMMS. Once loaded, no transformation or manipulation of the data is needed.

7.2 USER INTERFACE

In this section the user interface is discussed. In order to give a visualization, some pages are presented
in the appendix.

In the opening page (see Appendix A), the user can choose the excel-file that contains the data he
wants to analyze and base the cost estimation upon. Once loaded, on this page the user can select the
dependent and independent variable and directly see the simple linear regression line through a
scatterplot. Outliers in this scatterplot can easily be detected by displaying the standard deviation
lines and the outliers can easily be deleted.

Subsets can be selected, on different levels. Individual observations can be deselected, but whole
regions or countries can be selected and deselected as well. In this way, for every subset can be easily
seen if there are linear relations between variables. In addition, other basic information, like
correlation, the fit represented by R2 and R2-adjusted, number of observations and averages of the
selected subsets are shown.

In the second page the same selection panes as on the opening page are available, but in this case
multiple independent variables can be selected to evaluate the fit with the dependent variable. The
results of this multi-regression are shown, as well as averages of the selected subsets.

On the third page (see Appendix B), the selected data can be clustered. The number of clusters and
the cost driver on which the clustering is based can be selected. The observations are then clustered
and for each cluster the characteristics are shown: the fit represented by R2 and R2-adjusted, number
of observations and averages of the selected cluster are shown. There is also an option to let the
algorithm decide the number of clusters that provide the best fit.

On the last page the user can select an estimation method and put in the needed parameters and
easily obtain an estimation of the costs based on the input and the used method.

39
7.3 PERFORMANCE
The cost estimation equations must be determined before an estimation can be performed. In order
to perform these, the regression equations are all defined using the optimization engine of AIMMS.
The simple linear regression and the multi-regression are immediately solved to an optimal solution.
The nonlinear regression, the clustering algorithm and the extended regression are somewhat harder
to solve and therefore not solved to optimality, but to suboptimal solutions. The advantage of this is,
that the suboptimal solutions are acceptable and are determined within seconds. Other calculations
do not take notable time.

7.4 CONCLUSION
This section provides in answering the fifth research question:

5. How can the insights gained in this research be implemented?

The analysis, as well as the actual cost estimation is all integrated in AIMMS. It is designed in such a
way the user gets insight in the relations within the data and is able to select subsets and perform
multiple methods of analysis. The user does not have to do any calculations by itself and it can do the
analysis by going through the pages. At the last page the user can select the preferred, or the best,
cost estimation method to do the actual cost estimation.

40
8 CONCLUSIONS & DISCUSSION
Each chapter in this research the answer to a research question. The conclusion in Section 8.1 states
whether the problem statement of this research is solved. The problem statement is as follows:

“In order to make good cost estimations for newly built warehouses or depots, build a generic, user-
friendly tool that can quickly and accurately estimate the periodic costs of a warehouse”, with:

- Generic: Regardless of sector and the availability of data, the tool must be able to do accurate
estimations. Basic cost and operational data, such as the total costs and the amount of
products stored in or passing through a warehouse per period, can be expected from all
customers.
- Accurately: Given the situation (the availability of data), the tool must use the most
appropriate method to provide a reasonable estimation. The goal, as set by ORTEC, is to
perform cost estimations with a maximum deviation of 10% in 90% of the cases.
- Tool: The desired platform for this tool is AIMMS. The tool must be designed in such way it
can later be implemented within OSCD.
- Quickly: As part of an OSCD-study the tool must be fast, preferably providing an estimation
within the order of seconds.
- Costs: The total periodic operating costs of a warehouse.

Section 8.2 gives further recommendations based on this research. In Section 8.3 the limitations of
this research are discussed, followed by ideas for future research in Section 8.4.

8.1 CONCLUSION
This section provides in determining whether the problem statement of this research, as mentioned
above, is solved.

Based on interviews, scientific literature and the analysis of customer-data, the most important cost
drivers are identified:

- Throughput
- Building area
- Automation level
- Number of employees
- Country
- Area

of which throughput is the main cost driver used in this research. This has three reasons:

- The cost driver analysis resulted in throughput as being the most important cost driver.
- Throughput is typical input for a supply chain study, while building area is not.
- The most observations in the analyzed case contain throughput. This is not the case for the
other possible cost drivers.

After having set out the different kind of cost estimation methods available, a selection is made based
on the requirements as set in the problem statement and based on the data that is typically provided
by customers. The cost estimation methods that have been tested on estimation accuracy within the
case study are therefore:

 Simple linear regression

41
 o A linear regression equation, based on throughput as cost driver.
 Activity based costing
o Assigning all costs to individual products.
 Activity based costing per country
o Assigning all costs to individual products, per country.
 Clustering
The data is first clustered in such a way that the fit between observations within each cluster
is optimal.
 Multi-regression
The multi-regression has two cost drivers: throughput and building area.
 Extended regression
A regression model based on throughput and additional adjustment factors: country, area
and automation level.
 Normalized regression
Simple linear regression with normalized data; the costs are normalized using the labor cost
index of Eurostat (2015).
 Nonlinear regression
A standard linear regression model, extended with an exponential factor on the cost driver:
throughput.

The main cost driver used in all the cost estimation models is throughput and the estimator is in all
cases the total costs. The multi-regression is based on both throughput and building area and ABC
per country and extended regression include an additional country factor.

The cost estimation methods are all applied to the dataset of the running example in this research;
the packaging distributor. After analyzing the estimations from the different methods, three of them
produce the most accurate estimations, based on the sample standard deviation: nonlinear
regression, multi-regression and extended regression. When looking at the goal of ORTEC:

- Perform cost estimations with a maximum deviation of 10% in 90% of the cases

this objective is not met. The most accurate estimations are gathered by nonlinear regression, which
could, in the best case, perform cost estimations with a maximum deviation of 40% in 90% of the
cases.

The different methods are all integrated within AIMMS in a user-friendly environment. The
developed tool works intuitively and fast. It has ability to identify relationships within the data and
perform analysis. The different cost estimation methods can all be used.

When evaluating all the requirements as mentioned in the problem statement, the results are as
follows:

- Generic:
All methods described are highly generically applicable.
Objective met: Yes
- Accurately:
With the current set-up of the cost estimation methods.
Objective met: No
- Tool:
The analysis of data, as well as the cost estimation methods are all implemented in AIMMS, in
a user-friendly tool.

42
 Objective met: Yes
Quickly:
The estimations are all generated within the order of seconds.
Objective met: Yes
- Costs:
All methods focus on estimating the total periodic operating costs of a warehouse.
Objective met: Yes

This research resulted in identifying the most important cost drivers of a warehouse. After that, the
most suitable methods have been analyzed and modelled in such a way that cost estimations can be
performed. The estimations following from these methods are all analyzed based on accuracy and the
strongest estimation methods are selected.

All requirements as set by ORTEC are taken into account. The described methods are all suitable for
doing cost estimations, but in the current set-up they do not lead to the required accuracy.

8.2 RECOMMENDATIONS
Based on this research, additional recommendations are formulated:

- Gather datasets from customers with more cost drivers available and preferably more
observations. In this way, the analysis can be more thorough and more sophisticated cost
estimation models can be developed and evaluated. This will likely increase the accuracy of
the estimations.

- Collect multiple datasets of different customers within different industries. This research is
based on one individual client, so it would be interesting to see if the same conclusions can
be drawn over a combined dataset containing multiple clients and industries. In this way
general rules can be developed. Other methods, like machine learning could also be applicable
to these larger datasets.

8.3 DISCUSSION
After having drawn conclusions and recommendations regarding this research at ORTEC, this section
discusses the limitations of this research.

- As this research mainly focuses on warehouse cost estimation for supply chain studies, the
analysis performed is on a strategical level. Therefore, the data is all aggregated to a certain
period, in this case a one-year period. Better cost models and resulting estimations could be
obtained by taking into account additional, perhaps lower-level, characteristics.

- The case study performed in this research is based on one single customer that distributes
homogenous products. The findings within this case study are not guaranteed to be
applicable to other customers with different products or within a different sector.

- The data used in the case study is far from ideal. Not many cost drivers are available, so the
analysis is somewhat limited. Better results and additional insights could have been found
with a more extended dataset or multiple datasets.

43
8.4 FUTURE RESEARCH
Based on the findings in the conclusion, recommendation and discussion, this section discusses the
subjects of future research.

- As mentioned earlier in this chapter, more datasets with more cost drivers must be collected
in order to develop more accurate cost estimation models. Next to that, combined datasets
from different customers and sectors could provide general rules that are applicable to every
customer in every sector.

- In this research, only the total costs are taken into account. A distinction can be made between
several cost entries and what drives these. For this research, this was out of scope, but
investigating this may lead to more accurate cost estimations.

- For this research, a high-level approach is used, which did not result in the required accuracy
of the cost estimations. This choice is made because customers do typically not have all the
required data available or it is too costly to gain the needed data. Further research must be
done to find a good balance between the gathering of data and the power of the estimations.

44
BIBLIOGRAPHY
Anderberg, M. R. (2014). Cluster analysis for applications: probability and mathematical statistics: a
series of monographs and textbooks (Vol. 19). Academic press.

Angelis, D. I., & Lee, C. Y. (1996). Strategic investment analysis using activity based costing concepts
and analytical hierarchy process techniques. International Journal of Production Research, 34(5),
1331-1345.

Armstrong, J. S. (2001). Selecting forecasting methods. Principles of Forecasting, 365–386.

Bartholdi, J., & Hankman, S. (2011). Warehouse & distribution science 2007.

Breiman, L. (2001). Statistical Modeling: The Two Cultures. Statistical Science, 16(3), 199–215.

Cooper, R., & Kaplan, R. S. (1988). Measure Costs Right: Make the Right Decision. Harvard Business
Review, 66(5), 96–103.

Desarbo, W. S. (1989). A simulated annealing methodology for clusterwise linear regression. Society,
54(4), 707–736.

Desarbo, W. S., & Cron, W. L. (1988). A maximum likelihood methodology for clusterwise linear
regression. Journal of Classification, 5(2), 249–282.

Dysert, L. R. (2008). An Introduction to Parametric Estimating. AACE International Transactions, 1–8.

E. Backer and A. Jain (1981), “A clustering performance measure based on fuzzy set decomposition,”
IEEE Trans. Pattern Anal. Mach. Intell., vol. PAMI-3, no. 1, pp. 66–75, Jan. 1981

Everaert, P., Bruggeman, W., Sarens, G., Anderson, S. R., & Levant, Y. (2008). Cost modeling in logistics
using time driven ABC. International Journal of Physical Distribution & Logistics Management, 38(3),
172–191.

Fox, J. (2015). Applied regression analysis and generalized linear models. Sage Publications.

Friedman, J., Hastie, T., & Tibshirani, R. (2008). The elements of statistical learning (Vol. 2). Springer,
Berlin: Springer series in statistics.

Friedman, N., Geiger, D., & Goldszmidt, M. (1997). Bayesian network classifiers. Machine learning,
29(2-3), 131-163.

Galton, F. (1885). Regression Towards Mediocrity in Hereditary Stature. The Journal of the
Anthropological Institute of Great Britain and Ireland. Royal Anthropological Institute of Great Britain
and Ireland, 15(1886), 246–263.

GAO. (2009). GAO Cost Estimating and Assessment Guide. GAO Cost Estimating And Assessment
Guide, (March), 440.

Goh, M., Jihong, O., & Chung-Piaw, T. (2001). Warehouse sizing to minimize inventory and storage
costs. Naval Research Logistics, 48(4), 299–312.

Gong, Y., Zhang, Z., & Wang, S. (2009). Stochastic Modelling and Analysis of Warehouse Operations.

Hung, M. S. (1984). Economic sizing of warehouses: a linear programming approach, II(I), 13–18.

45
J. MacQueen (1967), “Some methods for classification and analysis of multivariate observations,” in
Proc. 5th Berkeley Symp., vol. 1, pp. 281–297.

Kotsiantis, S. B., Zaharakis, I., & Pintelas, P. (2007). Supervised machine learning: A review of
classification techniques.

Legendre, A. M. (1805). Nouvelles méthodes pour la détermination des orbites des comètes (No. 1).
F. Didot.

Michalski, R. S., Carbonell, J. G., & Mitchell, T. M. (Eds.). (2013). Machine learning: An artificial
intelligence approach. Springer Science & Business Media.

Montgomery, D. C., Peck, E. A., & Vining, G. G. (2015). Introduction to linear regression analysis. John
Wiley & Sons.

P. Sneath (1957), “The application of computers to taxonomy,” J. Gen. Micro- biol., vol. 17, pp. 201–
226.

Richards, G. (2010). Warehouse Management.

Robert L. Thorndike (1953). "Who Belongs in the Family?". Psychometrika 18 (4): 267–276

Späth, H. (1982). Algorithm 48: A fast algorithm for clusterwise linear regression. Computing, 29, 175-
181.

T. Sorensen (1948), “A method of establishing groups of equal amplitude in plant sociology based on
similarity of species content and its application to analyzes of the vegetation on Danish commons,”
Biologiske Skrifter, vol. 5, pp. 1–34.

Trost, S. M., Asce, M., Oberlender, G. D., & Asce, F. (2003). Predicting Accuracy of Early Cost Estimates
Using Factor Analysis and Multivariate Regression, 129(2), 198–204.

Varila, M., Seppänen, M., & Suomala, P. (2007). Detailed cost modelling: a case study in warehouse
logistics. International Journal of Physical Distribution & Logistics Management, 37(3), 184–200.

Xu, R. (2005). Survey of clustering algorithms for MANET. IEEE Transactions on Neural Networks, 16(3),
645–678.

Park, Y. H., & Webster, D. B. (1989). Modelling of three-dimensional warehouse systems. the
international journal of production research, 27(6), 985-1003.

46
APPENDIX A: OPENING PAGE

APPENDIX B: CLUSTERING

47
APPENDIX C: ESTIMATION PAGE

48