An Inventory Control System For Spare Parts at A Refiner

Production, Manufacturing and Logistics
An inventory control system for spare parts at a renery:

An empirical comparison of dierent re-order point methods
Eric Porras
a,
*
, Rommert Dekker
b
a
Instituto Tecnolo gico y de Estudios Superiores de Monterrey, Campus Santa Fe, 01389 Mexico City, Mexico
b
Econometric Institute, Erasmus University Rotterdam, P.O. Box 1738, 3000 DR Rotterdam, The Netherlands
Received 5 August 2005; accepted 9 November 2006
Abstract
Inventory control of spare parts is essential to many organizations, since excess inventory leads to high holding costs
and stock outs can have a great impact on operations performance. This paper compares dierent re-order point methods
for eective spare parts inventory control, motivated by a case study at a large oil renery. Dierent demand modeling
techniques and inventory policies are evaluated using real data.
2006 Elsevier B.V. All rights reserved.
Keywords: Spare parts; Inventory control; Re-order points; Demand classes; Service levels
1. Introduction
Eective inventory management of spare parts is
essential to many companies, from capital-intensive
manufacturers to service organizations, such as car
manufacturers, chemical plants, telecom companies
and airlines. Dierent from work-in-process (WIP)
and nished product inventories, which are driven
by production processes and customer demands,
spare parts are kept in stock to support maintenance
operations and to protect against equipment fail-
ures. Although this function is well understood
by maintenance managers, many companies face
the challenge of keeping on stock large inventories
of spares with excessive associated holding and
obsolescence costs. Thus, eective cost analysis can
be an important tool to evaluate the eects of stock
control decisions related to spare parts. However,
the diculty in assessing good strategies for the
management of spare parts lies in their specic nat-
ure, normally very slow-moving parts with highly
stochastic and erratic demands. For example, typi-
cal industrial data sets comprise limited demand his-
tory with long streams of zero demand values and a
few large demands (Willemain et al., 2004). This
makes the estimation of the lead time demand
(LTD) distributions very dicult, which is essential
to obtain the control parameters of most inventory
policies. Although dierent inventory models have
been proposed in the literature to tackle this prob-
lem (see next section), there is a lack of empirical
testing of theoretical models with data from real
industrial environments.
0377-2217/$ - see front matter 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2006.11.008
*
Corresponding author. Tel.: +52 55 91778000.
E-mail address: [email protected] (E. Porras).
European Journal of Operational Research xxx (2007) xxxxxx
www.elsevier.com/locate/ejor
ARTICLE IN PRESS
Please cite this article in press as: Porras, E., Dekker, R., An inventory control system for spare parts at a renery: ...,
Eur. J. Oper. Res. (2007), doi:10.1016/j.ejor.2006.11.008
This paper concerns a study on spare parts at a
major oil renery in the Netherlands, which con-
sisted of two phases. In the rst phase a case study
was conducted with the company with the idea to
optimize the use of the SAP system. As a result,
some improvement measures were provided and
later implemented by the company. The second
phase was focused on the analysis of the demand
data provided by the company. This paper reports
on the ndings related to this phase, where the
objective was to perform an empirical comparison
of dierent inventory models.
The aforementioned company keeps stock of a
large number of spare parts related to equipment
used in its petrochemical processes. Although these
stocks are essential for the continuity of its opera-
tions, management was concerned with the savings
opportunities at the process oor by having better
inventory control of its spare parts, whose value
was worth at the moment of the study more than
27 million euros. One major diculty of the study
was the limited demand history available.
By describing the case, we make general observa-
tions about the practical aspects of inventory con-
trol. Moreover, our aim is to compare various
policies with real demand data from the case to
see which one is best under what circumstances.
Common methods presented in the literature rather
use given statistical demand distributions to assess
the performance of inventory models. Conse-
quently, with our methodology we can better iden-
tify the real limitations of industrial data sets.
The remainder of the paper is organized as fol-
lows: the next section presents briey related litera-
ture. Section 3 includes the case study description.
Next the methodology is explained in Section 4.
The computation results are included in Section 5
and the nal conclusions are presented in the last
section.
2. Review of related literature
One of the major areas of inventory research over
the past decades is the one related to the manage-
ment of spare parts inventories. Although theoreti-
cal models for slow-moving items are abundant in
inventory literature since 1965, case studies are
few (for a comprehensive overview of recent litera-
ture on spare parts management see Kennedy
et al. (2002)). These studies are generally concen-
trated on the mathematical optimization of inven-
tory models but unlike our work they do not
validate them using empirical data.
In the arena of theoretical models, one of the
most extensively studied inventory policies is the
so-called (S 1, S) model, a particular case of
(s, S) models, with an underlying Poisson demand
distribution (see Feeney and Sherbrooke, 1966).
Although well studied and suitable for slow-moving
items, this type of policy requires continuous review
of the inventory system. Moreover, the Poisson dis-
tribution assumes randomness of demand, with
interarrival times between unit size demands follow-
ing an exponential distribution. This distribution
needs no information of demand other than the
average demand, which is the sole parameter of
the demand distribution. When transactions are lar-
ger than unit size, authors have proposed the use of
compound-Poisson models (see Williams, 1984; Sil-
ver et al., 1971). However, these models are more
dicult to apply in practice because they need an
assumption on the compounding distribution. For
example Williams (1984) developed a method to
identify sporadic demand items, where three param-
eters are needed: one for the exponential distribu-
tion of interarrival times of demands, and two
parameters of an underlying gamma distribution
for the demand size.
Most of the empirical studies in spare parts liter-
ature are focused on testing forecasting methods for
demand of slow-moving items rather than on imple-
menting inventory models. This is an important dis-
tinction since forecast methods are used to estimate
point forecast of the mean (like the moving average
method) while for evaluating control parameters of
inventory models (like the (s, Q) model) one needs
an estimation of the entire LTD distribution. More-
over, inventory models are used to meet specic cus-
tomer service levels in the long run, while
forecasting models aim to obtain accurate demand
forecasts as determined by the mean average per-
centage error (MAPE) or the mean square error
(MSE). In this area of research Ghobbar and Friend
(2003) present a comparative study of 13 dierent
forecasting methods for the management of spare
parts in the aviation industry. They use the MAPE
measurement applied to forecast errors to assess
the accuracy of the dierent methods but no inven-
tory models are included in the study. They conrm
the superiority of the weighted moving average and
Crostons methods over exponential smoothing and
seasonal regression models. Silver et al. (1998) also
advise on the use of Crostons method for products
2 E. Porras, R. Dekker / European Journal of Operational Research xxx (2007) xxxxxx
ARTICLE IN PRESS
with intermittent and erratic demand. This method
(Croston, 1972) assumes that the LTD has a normal
distribution, and estimates the mean demand per
period by applying exponential smoothing sepa-
rately to the intervals between non-zero demands
and their sizes.
Willemain et al. (2004) propose the use of a mod-
ied bootstrap method to forecast intermittent
demand of service parts, and they implement the
method on a large industrial data set. Crostons
method and exponential smoothing are evaluated
as well, yet like in the previous paper no attempt
is made to implement an inventory control model.
They show that the modied bootstrap method pro-
duces more accurate forecasts (based on the MAPE
measurement) than the exponential smoothing and
the Crostons method. We use in the study the boot-
strap method proposed by Willemain and we com-
pare it with the performance of an empirical
distribution model. We assess also the performance
of models based on Poisson and normal demand
distributions.
A similar research to the present study is pre-
sented in Strijbosch et al. (2000), where the perfor-
mance of two dierent (s, Q) models for spare
parts in a production plant environment is exam-
ined. Unlike our method, they test the inventory
models proposed using simulation where demands
are generated from an Erlang distribution, whereas
we assess the inventory models using the historical
demand data for the items (around 8000 items).
Also related to our study, Gelders and van Looy
(1978) presented a case study carried out in a large
petrochemical plant. They developed dierent
inventory models to control slow and fast moving
items, which were clustered in classes using ABC
analysis together with criticality and value consider-
ations. As they had limited information on con-
sumption rates for slow movers, a Poisson
underlying distribution was assumed to compare
between existing practices and the models proposed.
We use in our study a similar approach, but dier-
ent to their study we estimate the LTD distribution
using the methods mentioned above and we test the
models with real demand data, rather than using
simulation.
3. The case study
In this section we describe the problem setting,
the demand data, the classication criteria for the
parts and the costs.
3.1. Problem description
The company under study consists of a major
petrochemical complex located in the Netherlands,
which includes 60 dierent plants divided in chemi-
cals manufacturing and oil rening. The complex
dates from 1930, and many new installations have
been added since then. A large part of it, however,
stems from the 1960s. The procurement department
oers service to all plants. There is one central ware-
house owned by the company. At the moment of the
study (2000), there were in total 130 thousand cata-
logued materials, of which only 43 thousand were
kept on stock at the site, with a total value of more
than 27 million euros. There are 22 additional small
de-central storages on site, containing fast moving
materials that can be directly used if needed. No
stock registration is done for these items and they
are replenished on a batch basis. Therefore, we only
need to consider a single stock echelon, being the
warehouse as user of spares for equipment and
not a producer of parts. In total there are 180 thou-
sand requests of material per year, both for non-
stock and stock materials. Requests for materials
kept on stock are supplied from available stock. If
there is shortage of a material an emergency replen-
ishment order is generated.
Controlling 43 thousand materials represents a
dicult task, especially because of the dierences
in types and consumption patterns associated with
them. It also requires ecient use of the manpower
available and of the information system at hand.
Until 1997 an in-house developed information
system for inventory control was used by the com-
pany. In 1997 they moved to the information sys-
tem SAP R/3, which is a complete ERP-system,
but not specic for inventory control. Almost the
whole demand history before 1997 has been lost
in the transfer to SAP. Within SAP, the company
applied the MM (materials management) module
for the control of its spare parts. Since SAP
evolved out of MRP systems for the manufactur-
ing and assembly industry, the MM module is very
much based on the MRP planning philosophy (see
Heizer and Bender, 2001). Demand is expressed by
actual orders or by forecasts of demands. Next
demand of end products is converted to demand
for assemblies, components and parts. Stock con-
trol is performed in SAP on a periodic basis (so-
called periodic review). Items are ordered when
the MRP run is made. The SAP user can set the
appropriate time interval, e.g. daily, weekly or
E. Porras, R. Dekker / European Journal of Operational Research xxx (2007) xxxxxx 3
ARTICLE IN PRESS
monthly. At the company they run the MRP every
week.
The actual stock control within SAP occurs in
terms of minmax levels (equivalently to (s, S) poli-
cies), or MRP-type control based on lead times.
Minor functionality is available in SAP to deter-
mine the minimum level s and the maximum level
S. Safety stocks can be used to determine the re-
order level s, and lot sizing methods are available
to evaluate the dierence between s and S. Before
the project 90% of the control levels were set manu-
ally, and afterwards some 70%. As a result still
many replenishment orders were checked manually
before sending them out.
With respect to forecasting, several methods are
available in SAP, like exponential smoothing and
moving averages, both with trends and seasonality.
It is however the intermittent nature of demand that
makes the application of these methods particularly
dicult to spare parts. For the determination of
safety stock levels, the normal loss model is avail-
able which approximates the demand during the
lead time with a normal distribution. This model
works with the cycle service level as service level
objective. However, no ll rate service levels can
be dened within the MM module. A more striking
aspect of SAP is that within its functionality no con-
tinuous review models can be implemented. There-
fore the classical and much advised (S 1, S)
model with Poisson distributed demand over the
lead time cannot be applied.
3.2. Data structure
Statistical information for the consumption of
spare parts was available for 5 years (the last year
only until August). The demand information was
recorded in monthly periods, so a total of 55 periods
of demand information was available for the study.
One important limitation of the demand set was
that it did not specify whether demands were due
to failures or preventive maintenance activities.
Dierent parts used by the company are divided
in two main categories: materials related to a piece
of equipment and the ones not related to any
particular equipment, like protecting shoes, helmets,
general-purposed electrical equipment and instru-
mentation. From the total of 43,000 materials in
stock, 14,383 were spare parts, accounting for 80%
of the total stock value. These spare parts are the
focus of the present study. The parts related to
equipment are classied according to criticality
codes as follows:
High (H): Unavailability of these materials
would result in expensive downtime or cause dan-
ger to the safety of the people and the environ-
ment. Risk taken in the process of ordering and
stocking cannot be justied.
Medium (M): Unavailability of these materials
would result in signicant loss of production,
but does not endanger the safety of the people
or the environment. A calculated risk can be
taken in the process of ordering and stocking.
Low (L): Unavailability of these materials would
not result in serious eects on the processes or on
the safety of the people and the environment.
The previous classication is made on expert
judgment and no quantitative methods are used to
date. A further inspection of the materials with crit-
icality code leads to a more rened classication:
materials that are uniquely installed in a particular
piece of equipment (60% of the materials related to
equipment), and materials which are related to more
than one piece of equipment of dierent criticality
codes. That means that there are spare parts that
have combined criticality codes (H/M/L, M/L)
depending on whether they are installed in multiple
piece of equipment of dierent criticality. The com-
pany used these codes to decide on the stock levels
of the dierent parts. Thus, items identied as highly
critical should be on stock since they require high ll
rates, low critical ones destock, and medium critical
ones on stock depending on cost-eective consider-
ations. However, as no models are available in
SAP that incorporate criticality considerations,
these levels were set mostly by expert judgment.
3.2.1. Classication of parts
A more rened analysis of the spare parts data
revealed that important dierences among them
existed not only in terms of criticality codes but also
with respect to demand and price. Therefore, we
aimed at grouping them in dierent classes to see
whether we should apply dierent stock control
methods for dierent classes.
Below we describe the dierent classes considered
in this study.
3.2.1.1. Criticality classes. Based on the criticality
codes, the following criticality classes were dened
for the spare parts:
ARTICLE IN PRESS
Criticality class 1: H
Criticality class 2: H/M/L or H/L
Criticality class 3: M
Criticality class 4: M/L
Criticality class 5: L
Criticality class 6: Not related to any particular
piece of equipment
For the current policy used by the company, we
expect to observe that the service levels associated
with high critical items are higher than the ones
for low critical items (see discussion in Section 5.4).
3.2.1.2. Demand classes. The original data set con-
sisted of more than 14,000 spare parts, for which
we observed a high variability in demand patterns.
For example some parts had only 0/1 demands
while others experienced either few large demands
or no realization of demands during 5 years. For
other parts we observed large negative demands
due to returns. Thus, a classication was needed
for the spare parts based on consumption rates.
For parts with total positive demand over the ve-
year period and some demand values higher than
1, we identied from a histogram two main groups:
parts with relatively high total demand and parts
with low total demand. Although the boundary
between these two groups was not clearly identied
in the histogram of demand, from a Pareto analysis
we could reasonably establish it in 60 units. This
also implies that at least in one month demand is
at least 2. We observe that 90% of the items had a
demand below this value, and this contributed
25% of the total demand. At the same time, items
with 0/1 demands also had a total demand of less
than 60. According to this, we established the fol-
lowing demand classes for the parts:
Demand class 1: parts with only 0/1 demands.
Demand class 2: parts with total demand larger
than 0 but less than 60, and not only 0/1
demands.
Demand class 3: parts with total demand higher
than 60.
Demand class 4: parts with 1, 0, 1 demands.
Parts with negative total demand or parts with no
realizations of demand (all demands equal to zero)
were excluded from the analysis.
3.2.1.3. Price classes. For the spare parts in the data
set, ve dierent price levels were identied in a his-
togram. Table 1 shows the dierent price classes for
the spare parts. The parts recorded in SAP with a
price of 0 euros are items not owned by the materi-
als department (price class 1). We observed prices as
low as 0.01 euros for some parts (price class 2) and
the most expensive ones had a price of 20,000 euros
(price class 5).
3.2.1.4. Combined classes. Using the criticality,
demand and price classes, we include each item in
a combined class dened by three digits. Accord-
ingly, an item in class xyz corresponds to an item
with demand class x, criticality class y and price
class z. This classication allows us to optimize
the system per class rather than for individual items.
That is, once a service level is dened for the com-
bined class, the parameters for the dierent inven-
tory policies are evaluated for each item in the
class. Then a simulation tool is used to evaluate
the performance of the selected model of each indi-
vidual item using its demand data. Finally total
costs are aggregated across all items in the class.
In this way we aim at obtaining an optimization rule
for each combined class considered in the study (see
Section 4).
3.2.2. Considerations on item classes and anomalous
observations
The analysis of spare parts data is performed for
all combined classes incorporating demand classes
1, 2, 3 and 4. Thus, all criticality and price classes
which combine those demand classes are considered
for the evaluation of the inventory models, except
price class 1, since items with price zero do not have
associated holding costs. Since for items with total
negative demand a zero inventory policy is optimal,
we do not incorporate them in the analysis (4% of
the parts). Moreover, although negative demands
can be associated with returns due to preventive
Table 1
Price classes of parts
Price class 1 2 3 4 5
Price (p) in euros p = 0 0 < p 6 13.6 13.6 < p 6 169 169 < p 6 2112 p > 2112
Spare parts (total = 14383) 10% 19% 33% 29% 8%
ARTICLE IN PRESS
maintenance practices (parts that were ordered but
not actually installed) or with repaired parts that
were brought back to the system, we did not have
specic information in this respect. We also leave
out of the analysis items with no demand realiza-
tions in ve years (2.2% of the parts). From these
considerations we are left with 11,984 items. Addi-
tionally, we observed items with an error in the crit-
icality specication. These items accounted for 4.1%
of the total numbers of parts. After excluding these
items, we were nally left with 11,790 spare parts for
the analysis. We also identied in the data set one
particular month for which a large number of very
high demands was recorded. We considered that
this was due to an administrative rebooking in the
warehouse and thus we eliminated this month from
the data set in our analysis.
3.2.3. Lead times
The lead times for the spare parts were recorded
in days. However the demand data set for the items
was registered in months, without specication of
the day within the month that a particular demand
took place. Therefore, for ease of implementation in
the simulation we rounded the lead times o to full
months using 30 days per month. In this way an
item with a lead time of 80 days was considered to
have a lead time of 3 months. Observe that this con-
version is also necessary for the estimation of the
distribution of the lead time demand, since demand
forecasts for the items are produced in months.
Although the rounding up of lead times was intro-
duced to better cope with the demand data, by
doing this we take a conservative approach, in
which the service levels achieved by the system in
the simulation will be generally lower than they
are in reality.
3.3. Cost structure
In general, three types of costs are associated
with inventories: holding costs, ordering costs and
stockout costs. Holding costs represent the cost of
capital tied up in the spare parts inventory. An
annual xed rate of 25% was used in the study.
Ordering costs represent the cost associated with
placing an order for a spare part, which includes
the costs of telephone calls, inspection and handling
of the incoming items, paying the bill and registra-
tion of the parts. This cost is independent of the
number of parts included in the order. An ordering
cost of 36 euros was used in the study. Since our
objective is to evaluate the optimal balance between
service levels and holding costs we do not incorpo-
rate in this study stockout costs.
4. Methodology
We use two approaches for the optimization of
the spare parts inventory system under consider-
ation, namely an ex-ante and an ex-post approach.
In the ex-post procedure the same data set is used
for both tting and testing purposes. Opposite to
this, the ex-ante procedure, once a distribution has
been tted to the data, uses an entirely dierent
set for testing purposes. In this respect the ex-ante
approach is more appropriate from a scientic and
practical perspective, since in reality systems face
future unknown demands (Silver et al., 1998).
In order to achieve this, we divide the historical
demand data into two sets, namely a tting period
and a testing period. The tting period will be used
to estimate the lead time demand distribution
(LTD) which is used in turn to determine the inven-
tory policy parameters. The testing period is used to
perform a simulation to evaluate the performances
of the inventory policies selected and compare them
with the performance of the current one. We con-
sider two types of service levels, the cycle service
level (CSL) and the ll rate. The reason for using
the ex-post approach is that many industrial data
sets are rather short for forecasting purposes and
there is a lot of non-stationarity because of intro-
duction or outphasing of parts. This procedure will
give the advantage of using the whole data set to get
a better picture of the real demand process. The
performance of both approaches will be compared
to assess the validation conclusions made by each
one. We have to note however that in the data set
there are many items with only one or two demand
realizations, and therefore we expect highly variable
results.
One of the main issues we address is whether the-
oretical models can outperform stock analysts. In
practice this is dicult to assess because of lack of
information (e.g. short demand data sets, little
information on preventive maintenance practices),
as well as implementation constraints. For instance
real lead times of items are normally in days but in
the models one may prefer to use full periods of time
for ease of implementation. Other practical issues
like the performance of dierent methods to model
the demand process are explored as well in this
study. In order to achieve this, below we give the
ARTICLE IN PRESS
demand modeling methods used and next we
describe the inventory models considered.
4.1. Modeling the lead time demand
In inventory decision making, one needs to deter-
mine inventory control parameters, such as re-order
points and safety stocks. In order to do so, we need
a specication of the lead time demand distribution.
This is traditionally done by modeling the lead time
demand using common probability distributions
found in the literature, such as the normal distribu-
tion (Silver et al., 1998) or the Poisson distribution
(see Schultz, 1987). Other authors propose the use
of models based on forecast techniques such as
moving average or exponential smoothing (see Cro-
ston, 1972; Silver et al., 1998). In order to cope bet-
ter with real data sets and to give a more realistic
picture of demand, authors have proposed the use
of bootstrap techniques (Bookbinder and Lordahl,
1989; Efron and Tibshirani, 1993; Willemain et al.,
2004). As we are interested in the performance of
theoretical models using real data, we estimate the
LTD distribution using the Willemains bootstrap
method along with another novel procedure using
empirical data. In this way we estimate the distribu-
tion of demand over the lead time for each model,
which is used in turn to evaluate the parameters of
the inventory policies selected. However, dierent
to the methods found in the literature, our objective
is to use directly the real demand values observed to
assess the performance of the policies using the sim-
ulation tool. To keep the study tractable, we do not
apply updating for the estimation of the LTD distri-
bution. We also evaluate the performance of the sys-
tem using normal and Poisson distribution based
models. Below we describe these methods.
4.1.1. Willemains bootstrap method (W)
We implemented the modied bootstrap method
presented in Willemain et al. (2004). This method,
as compared to traditional bootstrap techniques pre-
sented in the literature, has the advantage of captur-
ing better the autocorrelations between demand
realizations, especially when dealing with intermit-
tent demands with a high proportion of zero values.
The method rst evaluates the empirical transition
probabilities between states of zero demand and
states of positive demand for the dierent items.
Then using this information, a stream of zero and
non-zero demands is randomly generated for a per-
iod of length equal to the lead time. The non-zero val-
ues are lled with demand values sample from the
data set. In this way estimates of LTD for each item
are obtained for a large number of realizations (1000
in this study). This information is nally used to esti-
mate the distribution of LTD. Willemain et al. (2004)
applied this method to nine large industrial data sets
of service parts inventories and compared it with the
exponential smoothing method and the Crostons
method. He concluded that the modied bootstrap
method gave the best performance of all three meth-
ods. In Fig. 1, we show a plot of an estimation of the
LTD distribution (CDF) using Willemains method
for an item corresponding to class 215. The item
(labelled #741) has a lead time of nine periods and
its demand data for the 55 periods is as follows: in
periods 5 and 17 it observed positive demands of unit
size each, and in period 9 a demand of 2. The rest of
the periods no demands were observed. Notice that
although only lead time demand values of 0, 1, 2
and 3 were realized in the data set, the method is able
to produce a LTD estimation where many other lead
time demand values are taken into account.
Once the LTD distribution is obtained, we can
use it to determine a re-order point to achieve a
given ll rate b as follows:
From the CDF, F(x), of LTD obtain the list of
possible re-order point values s, by setting s = x,
where x are the lead time demand values, and their
corresponding probabilities f(x). Now choose the
smallest s satisfying:
100b% 6 1
ESs
Q

100;
0 2 4 6 8 10
0.4
0.5
0.6
0.7
0.8
0.9
1
x
F
(
x
)
Item# 741 Willemain's CDF
Fig. 1. Cumulative distribution function using Willemains
method.
ARTICLE IN PRESS
where Q is a pre-determined lot size and ES(s) is the
expected units short for a given re-order point s,
which is evaluated as follows:
ESs
xjx>s
x sf x:
Note: The term 1 ES(s)/Q is the calculated ll rate
for a given s.
We use the above procedure to evaluate a re-
order point s for item #741 for a given ll rate of
82%. This item has a price of 4172 euros. Using this
information and the procedure describe in Section
4.2 we obtain Q = 1. Now following the above pro-
cedure we have: for s = 3 we obtain ES(s) = 0.254
with associated ll rate of 74.6%, and for s = 4 we
obtain ES(s) = 0.136 with associated ll rate of
86.4%. Thus, in this case s = 4 is the re-order point
value sought.
4.1.2. Empirical distribution of lead time demand (E)
We implemented an empirical model to estimate
the distribution of LTD. Dierently from the tradi-
tional bootstrap method, we construct a histogram
of demands over the lead time without sampling.
This method is new to the literature as no attempts
have been made to use it for inventory control.
Since demands are taken directly from the data set
over xed periods of time equal to the lead time, this
method also captures autocorrelations and xed
demand intervals due to preventive maintenance,
and is far easier to implement than the modied
bootstrap method described above. Using the
empirical model, we build once more the CDF for
the LTD of item #741 which is depicted in Fig. 2.
As the empirical model uses only the demand values
from the demand data set to construct the CDF,
fewer lead time demand values are produced in this
CDF than in the corresponding one using the Wille-
mains method (see Fig. 1). Using a similar proce-
dure as the one described for the Willemains
bootstrap method, we can use the empirical CDF
to determine a re-order point for a given ll rate.
As an illustration of this, using the empirical CDF
of item #741 we obtain for s = 1 and s = 2 calcu-
lated ll rates of 68.1% and 87.2%. Thus, for a given
ll rate of 82% we determine a re-order point value
s = 2.
4.1.3. Normal distribution (N)
We implement a normal based model assuming
that demand for the parts follows a normal distribu-
tion. To this end, the average (
D and standard devi-

ation (SD) of the observed period demand is
evaluated to estimate the parameters l
LTD
and r
LTD
of a normal LTD distribution, as follows:
l
LTD
D L
r
LTD
SD
L
p
;
where L is the lead time of an item in full periods of
time (days, months, etc.), and with D and SD eval-
uated using the whole data set of demands, includ-
ing zero and negative values. Thus, for integration
over a normal LTD distribution we use common
formulas found in the literature (e.g. Silver et al.,
1998 or Chopra and Meindl, 2004).
The normal distribution is not generally advised
for modeling the demand of slow-moving items,
for which a Poisson distribution is better recom-
mended (Silver et al., 1998). Thus, we do not expect
the normal based model to give better results than
the others. However, we want to investigate its per-
formance as compared to the other models consid-
ered. To this end, when we evaluate re-order
points based on the normal LTD distribution to
meet desired ll rates, the values obtained are
rounded up to integer values, and negative values
are set to zero. We require this since for our system
re-order points are dened as positive integer values
in accordance with the discrete demand for parts. In
the case of negative values which are set to zero no
compensation for the gain mass is applied in the
integration of the normal LTD distribution. This
causes only a minor distortion in our results as we
normally look at high service levels which have
associated positive re-order point values. We neither
correct for the gain mass associated with the round-
ing o of fractional values, as this has also a minor
eect. Accordingly, to determine a re-order point s
Item # 741: Empirical CDF
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
0 1 2 3 4
Lead time demand
Fig. 2. Empirical cumulative distribution of lead time demand.
ARTICLE IN PRESS
for a given ll rate b, we use a similar procedure as
the Excel goal seek routine (see Chopra and
Meindl, 2004), by rst nding a z-value satisfying:
100b% 6 1
r
LTD
UNLIz
Q

100;
where UNLI(z) is the unit normal loss integral asso-
ciated with the unit normal variate z. Then s follows
from s = l
LTD
+ z r
LTD
, rounded up to the nearest
integer. Recall that z corresponds to a re-order
point s associated with a CSL value (see Silver
et al., 1998). Thus the product in the denominator
of the above formula gives the expected units short
for a given re-order level s and the whole right-hand
term gives the calculated ll rate. As mentioned in
Silver et al. (1998), the previous formula underesti-
mates the true ll rate if r
LTD
is large relative to
Q. Therefore, a correction should be made in the
numerator substituting the term r
LTD
UNLI(z)
by r
LTD
(UNLI(z) UNLI(z + Q/r
LTD
)). By using
the uncorrected formula, we obtained conservative
values of the ll rates, which did not have a major
impact in the optimization of the system (see results
in Sections 5.1 and 5.2).
Using the data for item #741 we obtain l
LTD
=
0.654 and r
LTD
= 0.975. Now, applying the above
procedure for a given ll rate of 82% we have:
For s = 1: CSL = 63.8%, z = 0.354, UNLI(z) =
0.247, calculated ll rate = 75.9%.
For s = 2: CSL = 91.6%, z = 1.379, UNLI(z) =
0.038, calculated ll rate = 96.3%.
Thus, s = 2 is the re-order point to be selected.
Remark 1. Although in many situations the nor-
mality assumption is not satised, this distribution
has widely been used in practice and it is available in
many packages. This is due to the simplicity to
evaluate re-order points and other parameters based
on the normal distribution.
4.1.4. Poisson distribution (P)
Silver et al. (1998) suggest that the Poisson distri-
bution is suitable to model demand of slow-moving
items. We use the Poisson distribution to estimate
the LTD distribution for items in demand class 1.
The reason is that demands for these items are of
unit size, and hence the basic assumption of the
Poisson distribution is satised. The only parameter
of the Poisson distribution, the average rate of
demand over the lead time, is estimated from the
demand data for the dierent items. For other
demand classes, a compound-Poisson based model
would be more appropriate, but it is not clear
beforehand which batch distribution should be used
as often very specic values are observed (e.g. item
15 has demands of 30, 80 and 120 on a total of
792 installed). The Poisson-based model is com-
pared to the normal, empirical and bootstrap meth-
ods for demand class 1.
In order to determine a re-order point for a given
ll rate using the Poisson model, we can apply a
similar procedure as the one described above for
the Willemains or the empirical method, substitut-
ing f(x) for the Poisson probabilities (with parame-
ter k
LTD
D L and using s = x as the set of
positive integers 1, 2, . . . For the Poisson model we
use a (s, S)-type policy with s = S 1 (see Section
4.2). Applying the above procedure to the data of
item #741 and a given ll rate of 82% we obtain:
For s = 0: calculated ll rate = 34.5%.
Thus, the re-order point value selected is s = 1.
Note: As for item #741 we observed demand val-
ues higher than one, this item belongs to a demand
class where the Poisson model is not applied. There-
fore, the above calculation is presented for illustra-
tive purposes (see Table 2 in Section 4.2).
4.2. Inventory models
We use an (s, nQ) inventory policy for the system,
with the re-order point s evaluated using the LTD
distribution according to the modeling methods
described above. Thus, when overshooting of the
re-order point s cannot be overcome by the lot size
Q, an alternative lot size equal to nQ is ordered,
such that the inventory position is brought above
s, where n is an integer value. This is a common
practice in inventory management (see Silver et al.,
1998). The lot size Q will be evaluated according
to the economic order quantity (EOQ) using aver-
Table 2
Inventory models considered
Model Parameters Demand classes
Current policy (C) minmax (s, S) 1, 2, 3, 4
Poisson based model (P) (S 1, S) 1
Normal based model (N) (s
N
, nQ) 1, 2, 3, 4
Empirical based model (E) (s
E
, nQ) 1, 2, 3, 4
Willemain based model (W) (s
W
, nQ) 1, 2, 3, 4
ARTICLE IN PRESS
age annual demand. We round o the EOQ calcula-
tion according to Axsater (2000), as follows:
1. Evaluate: m bEOQc
2. Set Q
1 if m 0
m if m 6 0 and
EOQ
m
6
m1
EOQ
m 1 otherwise
When we use the Poisson distribution to estimate

the LTD, we use Q = 1. This model is often referred
to as (S 1, S) model, with s = S 1. As average
demand is generally low in this case, the EOQ calcu-
lation also yields a value of 1.
For the classes under study the proposed policies
are compared to the current (min, max) policy in
terms of the selected service level and total costs.
The aim of the proposed methodology is to optimize
the system, that is, to establish which model and
policy perform best under which conditions (see
Table 2).
4.2.1. Handling of large demands
Consider the situation in which an item with a
relatively short lead time observed unusual high
demands. For such an item the analysis becomes
dicult as the associated re-order point (say evalu-
ated according to a normal LTD distribution) is
likely to be overshot when using a simulation tool
to assess the performance of a given inventory
model. To illustrate this we give in Table 3 the
demand data and other relevant information for
item #1307, which has a lead time of one month.
For this item we consider that the demand value
of 450 is an outlier, since this value is larger than
the average of the rest of the positive demands plus
10 times their standard deviation. So here the aim is
to construct a tool that lters out this large demand.
To this end, we implement a demand lter in the fol-
lowing way: all demands larger than the average
plus k standard deviations (evaluated using only
positive demands) are ltered out. For this particu-
lar item observe that k = 3 will not produce the
desired result and hence a value k = 2 is a better
choice. In a similar way, for items belonging to
demand classes 2 and 3 we found that actually
k = 2 was the best selection for the demand lter.
As a result of applying the mentioned lter, 8.7%
of the positive demand values for these two classes
were ltered out.
Remark 2. Although for the optimization of the
system we used the above lter of demands, we also
assessed the eect of having large demands included
in the optimization process, for which we obtained
similar results (see Porras, 2005).
Remark 3. In the inventory analysis of all four
models we neglect the so-called undershoot. This
is the amount by which the inventory positions
drops below the re-order point. If the demand size
is always one, the re-order level is always exactly
reached, but if the order size is large, there can be
substantial drop below the re-order level. When cal-
culating the leadtime demand distribution, one
should take the undershoot distribution therefore
into account. We did not include it in our analysis
because we removed most large demands.
4.2.1.1. Remark on the implementation of the ex-post
and ex-ante approaches. Due to the limitations
inherent to the data set used in our simulation
study, it may well be that not enough information
is available for tting purposes. Therefore, items
with only zero or one positive demand during the
tting period are excluded from the analysis. This
consideration is used in both the ex-post and the
ex-ante approaches. As for classes with few items
(6 or less), they were excluded from the analysis in
both the ex-ante and the ex-post approach. Accord-
ing to these considerations, of the original 11,790
items, a total of 8494 were included in the ex-post
approach and 4326 in the ex-ante approach (see
analysis of results in Sections 5.1 and 5.2).
4.2.1.2. Remark on classication of items. We
include each item in a class according to the demand
Table 3
Demand data for item #1307
Demand value Number of
occurrences
Total
demand
0 47 0
87 1 87
100 4 400
120 1 120
150 1 150
450 1 450
Total 55 1207
Average of demands > 0 150.9
Standard deviation of
demands > 0
122.4
Average of positive
demands < 450
108.1
Standard deviation of positive
demands < 450
20.8
ARTICLE IN PRESS
pattern and other relevant information (criticality
and price). We do the classication using the whole
data set in both the ex-ante and the ex-post
approaches, since we want to assure that an item
in a certain category will not exhibit demand values
not corresponding to that category during the simu-
lation in the testing period.
4.3. Optimization of the system: A decomposition
approach
We focus on optimization of the system based on
service levels, for which both the ll rate and the
cycle service level are used. The utilization of service
levels to set safety stocks is a preferred method in
industry as opposed to cost minimization. The rea-
son for this is that the latter requires the evaluation
of stock out costs which depend on down time pen-
alties and other factors which are dicult to evalu-
ate in practice.
Considering the size of the system related to the
present study (recall that originally it comprises
some 14,000 items), we use a decomposition
approach for the optimization of the system. This
type of approach is well known to the literature,
where it is often used for the analysis of complex
systems, e.g. for the optimization of multi-echelon
inventory systems (see van der Heijden et al.,
1997). In our case the basic idea is to optimize the
system at group level, dened by the classes
described earlier, rather than doing it at item level.
Thus, we evaluate the dierent inventory parame-
ters for the items based on a single target service
level for all items in the class. Notice that deviations
of the realized service levels with respect to the tar-
get values are expected due to the discrete nature of
inventory levels. Thus, the objective of the method
is to nd the right level of a target service level that
optimizes the classes under study.
We rst introduce the following notation:
j index for each item j in class p, j = 1, . . . ,
N
p
, where N
p
is the size of class p
X identier for the model applied in the simu-
lation, where X = current policy (C), Pois-
son (P), Normal (N), Empirical (E) or
Willemain (W).
c
X
j
number of inventory cycles completed by
item j during the testing period using model
X. Note: an inventory cycle is dened be-
tween the placing of an order and its arrival
to the system, i.e. the inventory cycle over a
lead time (also referred to as replenishment
cycle)
soc
X
j
number of inventory cycles for item j with
stockouts during the testing period using
model X
D
t
j
total number of units demanded of item j
during the testing period
S
X
j
total number of units of item j supplied
from on hand stock over the testing period
using model X
CSL
X
p
cycle service level achieved by the system
using the model X
b
X
p
ll rate achieved by the system using model
X
b
p
target ll rate for class p
ES
j
(s
j
) expected units short for item j for a given
re-order point s
j
, according to the corre-
sponding LTD distribution
Now for each combined class p(xyz) we proceed
as follows:
1. Evaluate the service levels (CSL and ll rate)
achieved by the current system over the testing
period according to the ex-post or the ex-ante
approaches. These are the current service levels
of class p, which are dened by
CSL
C
p
1
N
p
j1
soc
C
j
N
p
j1
c
C
j
; b
C
p

N
p
j1
S
C
j
N
p
j1
D
t
j
:
2. For each item j, estimate the cumulative distribu-
tion of LTD over the tting period, according to
the model selected (see Section 4.1). Use b
p
b
C
p
together with the LTD distribution to estimate
the corresponding parameters of the inventory
policy (i.e. re-order point and lot size), in the fol-
lowing way:
Calculate Q
j
, the lot size, according to the
EOQ formula and round it o as explain in
Section 4.2. If class p = 1yz (demand class 1)
then set Q = 1.
Chose the smallest s
j
which satises a 100b
p
%
ll rate, i.e.:
100b
p
% 6 1
ES
j
s
j
Q
j

100:
3. For each item j, run the simulation over the test-
ing period applying the (s, nQ) or the (S 1, S)
ARTICLE IN PRESS
policy as required, using the parameters selected
in the previous point. For each model selected,
evaluate for class p the realized ll rate (rb
p
),
and its total costs (TC
p
), according to
rb
X
p

N
p
j1
S
X
j
N
p
j1
D
t
j
; TC
X
p;target

N
p
j1
TC
j;target
;
where TC
j,target
is the total cost of item j for the
target service level selected, which comprises the
holding and ordering costs. Compute the total
savings achieved by the selected model with re-
spect to the current system according to:
Total savings
model
p
TC
current
p;target
TC
model
p;target
:
Within the classes we expect deviations of the
realized service levels for the individual items.
In order to assess the magnitude of those devia-
tions, we evaluate the weighted variances of the
realized ll rate as follows:
Varrb
X
p

1
N
p
j1
c
X
j
N
p
j1
c
X
j
b
X
2
j
rb
X
p

2
:
An estimation of the standard deviation of the
target ll rate is thus given by
^ r
X
p

Varrb
X
p

:
4. For each model selected check whether the real-
ized service levels are better than the service levels
of the current system. If the answer is yes, calcu-
late total costs.
5. For classes with lower realized service levels or
negative savings, construct exchange service
level-cost curves for the dierent inventory poli-
cies, in the following way: dene a set of xed tar-
get values for the ll rate (b
p
), say from 1% to
100% in step sizes of 1%. Next evaluate the
parameters of the selected inventory model as in
point 2 (except for the current system). For each
value of the target service level obtain the corre-
sponding rb
p
, TC
p,target
and total savings. Finally,
identify the value of b
p
, for which the correspond-
ing realized ll rate equals or exceeds the current
ll rate (within 1% precision), and that has the
largest positive savings. In this way we optimize
the system among the dierent classes.
Remark 4. The previous method assumes that the
system is optimized using the ll rate as optimiza-
tion criterion, which plays the role of control
parameter. However, it is also possible to implement
the above procedure using the CSL for the evalu-
ation of re-order points in steps 2 and 5, in the
classical way. Even in the case for which the true
optimizing criterion is the ll rate, one may decide
to apply the CSL as control parameter due to
limitations associated with the software available,
which may not allow the evaluation of the ll rate.
Remark 5. When more than one order is outstand-
ing for an item in the simulation, that is when 2 or
more replenishments cycles overlap, we associate a
stockout with the immediately precedent cycle that
has not nished yet, thus causing only that cycle to
be a stockout cycle. This situation is not considered
in inventory text books such as Silver et al. (1998),
thus confusion as to how to register stockout cycles
may arise. To illustrate this consider the item
#2248 (see Fig. 3) with lot size of 2 units, lead time
of 2 periods and re-order point of 2 units. In period
19 the inventory position reaches the re-order point,
triggering the rst replenishment cycle. This cycle
ends at the beginning of period 21 when the system
is replenished and causing the net stock to reach 4
units. This cycle ends up with no stockouts (thus col-
oured white). The second cycle starts in period 22
with net stock still positive. Notice that before this
cycle ends, a third cycle starts in period 23, where
the net stock drops to 1. This causes the second
cycle to be a stockout cycle (thus coloured grey). Sim-
ilarly, the demand of 8 units in period 24 causes the
third cycle to be a stockout cycle, triggering a fourth
cycle. Here the question arises as to whether the
fourth cycle should be counted as stockout cycle or
not. Although these two last cycles overlap we only
consider the former to be a stockout cycle (one orders
after the demand occurrence). Notice that actually
the fourth cycle ends up with a positive net stock.
In this way and assuming no further demands after
period 31, the itemwill achieve a CSL of 60%. Notice
that if the fourth cycle was considered a stockout
cycle the CSL calculation would give 40%.
4.3.1. Remark on the implementation of the
optimization methodology
In the study we optimized the system based on
the ll rate, since this service level is of more practi-
cal importance for spare parts. We investigated
however the realized cycle service levels (CSL)
achieved by the dierent classes under the models
considered. We observed that in many cases for a
ARTICLE IN PRESS
particular target ll rate the corresponding CSL was
way below the current CSL. Therefore, in order to
prevent such low values of the CSL we imposed
an extra condition in Step 5 of the previous algo-
rithm. Accordingly, for the identication of the
optimal target ll rate we checked that the realized
CSL was within 5% of the current CSL as long as
the system still achieved positive savings. In some
cases (especially for the empirical and Willemains
models) this condition was dicult to satisfy exactly
for particular classes (see e.g. classes 213, 214 and
215 in Table 11), and hence the realized CSL was
below the 5% value. For the evaluation of the real-
ized CSL of class p for the dierent models we use
the following formula:
rCSL
X
p
1
N
p
j1
soc
X
j
N
p
j1
c
X
j
with corresponding variance given by
VarrCSL
X
p

1
N
p
j1
c
X
j
N
p
j1
c
X
j
1
soc
X
j
c
X
j

2
rCSL
X
p

2
:
4.3.2. Accuracy of the LTD modeling methods
Since we are interested in the performance of the
modeling methods applied to inventory policies, a
useful measure of the accuracy of the dierent mod-
els can be provided by the average of the absolute
dierence between the target and the realized service
levels. As we observed earlier, dierences are
expected due to the discreteness of the inventory lev-
els. We call this measure MAD, dened by
MAD
p
CSL

1
N
p
N
p
i1
jrCSL
i
a
p
j
for a target cycle service level a
p
; and
MAD
p
b

1
N
p
N
p
i1
jrb
i
b
p
j
for a target fill rate b
p
:
4.4. The simulation model
We implemented the proposed methodology in
MatLab v.7.0. The system was simulated over the
testing period using the dierent inventory models.
Recall that demands for the items in each class were
drawn directly from the data set and used as input
to the simulation. Thus, we were able to evaluate
the performance of the dierent policies and models
in a real inventory environment. This allowed us to
capture the peculiarities of the system under study
and to discern under which conditions which policy
performs better.
Since demands cannot occur with a higher fre-
quency of one per month, we consider that the sys-
tem is reviewed on a continuous basis, and therefore
valid for the application of the (s, nQ) and (S 1, S)
policies.
For the evaluation of inventory cycles and com-
putation of inventory related costs we use the classi-
cal inventory methodology. Thus, the (s, nQ) policy
is implemented in the classical way as explained in
Section 4.2.
4.4.1. Considerations on starting stocks
Since the system orders nQ units whenever the
inventory position drops below the re-order point
(or Q when it is at the re-order point), the maximum
level that the system can have at any point in time is
given by its re-order point s plus Q. Therefore, for
comparison purposes it is fair to consider the start-
ing stock of an item as its re-order level s plus its lot
size Q. For the current system, an equivalent choice
is to set the starting stock of an item at its max level,
Item # 2248
ROP = 2
Q = 2
L = 2 periods
Period 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
Demands 0 1 0 0 0 0 0 1 0 0 2 3 8 0 1 0 1 0 0 0
Net stock 4 3 3 3 3 3 3 2 2 4 2 -1 -7 -4 3 3 2 2 4 4
Inventory Position 4 3 3 3 3 3 3 4 4 4 4 4 4 4 3 3 4 4 4 4
1 2 5
Cycles 3
4 ?
Fig. 3. Example to illustrate the occurrence of stockout cycles.
ARTICLE IN PRESS
since this can be considered as an order-up-to level
S with S = s + Q. This choice of starting stocks
allows us to compare the performance of the dier-
ent inventory policies considered. Note that for an
innite horizon, the CSL is indierent to the choice
of starting stock, whereas the ll rate is not. How-
ever for a xed data set, the starting stock may also
have an eect on the cycle service level of an item,
since for higher starting stocks fewer inventory
cycles will be completed. As we use the same deni-
tion of starting stocks for all models, the total num-
ber of inventory cycles will be the same among
them, except for the current system. To see this
notice that the number of inventory cycles for an
item is independent of the choice of its re-order
point. Since for all models we use the same demand
values and Q is xed, it follows that the number of
cycles will be the same regardless of the model used.
5. Results from the simulation optimization
After having applied the classication criteria
explained in Section 3, and using the methodology
presented in the previous section, we performed an
optimization of the inventory system of spare parts.
In order to have a clear picture of the whole process,
we present in Fig. 4 a ow chart diagram of the
entire procedure, from cleaning and classication
of demand, price and criticality data for spare parts
until the optimization of the system itself.
In Tables 4 and 5, we give a summary of the
numerical results obtained from the simulation opti-
mization for the ex-post and ex-ante approaches,
using as optimization criterion the ll rate (for a
complete set of results of the dierent item classes
see Tables 11 and 12 in the Appendix). We include
the relevant classes under study, for which the real-
ized CSL, the realized ll rate and the total savings
are reported for each model. We also include the
total cost of the current system with the current
achieved service levels (target CSL and target ll
rate). For convenience, we use the following nota-
tion in the optimization results tables:
tCSL target cycle service level, that is, the cycle
service level achieved by the current system
(C) using the (min, max) policy
t ll rate target ll rate, that is, the ll rate achieved by
the current systemunder the (min, max) policy
rCSL realized cycle service level, that is the cycle
service level achieved by each of the four
inventory models under consideration,
these are: Normal (N), Poisson (P), Empir-
ical (E) and Willemain model (W)
Fig. 4. Flow chart diagram of the optimization study.
ARTICLE IN PRESS
Table 4
Summary of optimization results for the system under ex-post approach
Item
class
# of
items
Current system (C) Normal model (N) Poisson model (P) Empirical model (E) Willemain model (W)
Total
costs
a
Service level (%) Savings Service level (%) Savings
a
Service level (%) Savings Service level (%) Savings Service level (%)
tCSL t ll rate (N) vs.
(C)
rCSL r ll rate (P) vs.
(C)
rCSL r ll rate (E) vs.
(C)
rCSL r ll rate (W) vs.
(C)
rCSL r ll rate
Summary by demand class
1 2226 3,693,400 95.6 95.6 774,220 96.2 96.8 703,250 96.8 96.8 762,330 96.3 96.9 751,500 96.0 96.6
2 5185 9,469,700 84.3 86.1 408,750 81.1 90.7 202,860 80.0 91.7 170,000 79.1 89.7
3 608 685,920 77.2 85.7 31,370 70.0 90.2 15,411 37.4 81.4 17,797 37.8 81.7
4 475 1,676,100 96.5 96.6 92,713 96.3 97.3 104,950 96.3 97.3 61,126 96.7 97.5
Summary by criticality class
1 493 1,704,700 88.9 88.3 20,428 87.3 91.8 3541 92.9 92.9 5074 76.1 87.7 16,583 79.5 88.7
2 447 277,250 94.1 95.3 37,882 95.0 97.1 1215 99.8 99.8 28,617 91.2 97.5 21,805 89.8 96.4
3 3190 9,165,100 90.1 89.8 853,480 88.4 92.7 648,930 97.5 97.5 802,150 86.9 93.5 743,200 86.6 92.2
4 1395 905,370 92.6 92.9 107,170 90.7 94.9 20,420 97.2 97.0 83,448 90.0 95.7 54,293 89.9 94.2
5 372 634,320 90.8 89.8 61,002 90.4 93.4 1060 96.1 96.1 30,051 88.6 94.0 54,554 88.0 92.9
6 2597 2,838,300 79.5 85.4 227,090 75.3 90.6 28,077 95.7 95.6 115,530 70.8 89.9 74,396 68.8 88.2
Summary by price class
2 2223 524,350 83.2 90.0 179,030 79.4 97.0 5415 97.9 97.9 166,870 75.1 95.7 171,200 73.7 95.0
3 3266 1,608,000 87.9 88.5 245,180 84.7 91.3 25,627 98.2 98.2 115,240 83.3 92.7 73,012 83.3 91.4
4 2487 3,637,300 90.2 89.4 200,410 90.1 90.7 54,393 95.5 95.4 71,791 86.7 90.4 43,397 85.6 88.4
5 518 9,755,400 90.4 88.2 682,440 88.4 91.5 628,640 96.0 96.0 700,820 83.2 89.9 677,220 84.0 89.3
Total 8494 15,525,060 87.5 89.2 1,307,055 85.1 92.6 703,246 96.8 96.8 1,054,723 82.2 92.7 964,832 81.5 91.4
a
All cost and savings gures are in euros.
E
.
P
o
r
r
a
s
,
R
.
D
e
k
k
e
r
/
E
u
r
o
p
e
a
n
J
o
u
r
n
a
l
o
f
O
p
e
r
a
t
i
o
n
a
l
R
e
s
e
a
r
c
h
x
x
x
(
2
0
0
7
)
x
x
x
x
x
x
1
5
A
R
T
I
C
L
E
I
N
P
R
E
S
S
P
l
e
a
s
e
c
i
t
e
t
h
i
s
a
r
t
i
c
l
e
i
n
p
r
e
s
s
a
s
:
P
o
r
r
a
s
,
E
.
,
D
e
k
k
e
r
,
R
.
,
A
n
i
n
v
e
n
t
o
r
y
c
o
n
t
r
o
l
s
y
s
t
e
m
f
o
r
s
p
a
r
e
p
a
r
t
s
a
t
a
r
e
n
e
r
y
:
.
.
.
,
E
u
r
.
J
.
O
p
e
r
.
R
e
s
.
(
2
0
0
7
)
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
j
o
r
.
2
0
0
6
.
1
1
.
0
0
8
Table 5
Summary of optimization results for the system under ex-ante approach
Item
class
# of
items
Total
costs
a
Service level (%) Savings
a
Service level (%) Savings Service level (%) Savings Service level (%) Savings Service level (%)
tCSL t ll rate (N) vs.
(C)
rCSL r ll rate (P) vs.
(C)
rCSL r ll rate (E) vs.
(C)
rCSL r ll rate (W) vs.
(C)
rCSL r ll rate
1 800 504,650 95.6 95.7 94,010 96.5 97.2 63,575 96.3 96.3 96,409 96.2 97.1 36,831 96.1 97.2
2 2873 1,975,100 84.8 87.3 53,384 81.3 91.5 10,900 79.1 91.1 18,623 79.6 89.6
3 475 152,750 76.3 90.4 11,873 80.1 92.9 16,379 67.2 90.1 12,119 72.6 90.8
4 178 246,130 98.9 99.0 1816 96.6 97.6 2571 96.6 97.6 3110 96.6 97.6
1 132 152,500 86.5 85.7 540 80.7 90.7 468 87.0 87.0 14,322 87.1 91.5 9752 80.3 88.9
2 301 81,914 94.8 96.5 12,877 97.2 97.8 1496 100.0 100.0 14,668 85.1 96.6 8338 94.6 96.9
3 1457 1,701,000 90.0 90.5 51,385 85.6 91.7 56,230 96.5 96.5 63,253 83.1 91.1 18,236 82.7 89.9
4 818 212,440 93.3 94.3 29,042 93.4 96.1 4,659 97.7 97.7 26,538 92.6 95.7 22,704 91.7 94.8
5 159 84,877 94.8 95.2 6130 92.5 95.2 2398 98.7 98.7 2132 92.5 95.6 2947 90.8 94.7
6 1459 645,930 76.3 84.6 58,558 75.6 91.4 8578 94.0 94.0 28,848 71.8 90.5 21,215 74.2 89.9
2 1372 144,420 81.1 90.5 58,581 82.9 98.3 1919 95.2 95.2 63,142 77.6 97.8 58,886 84.3 98.3
3 1648 377,020 87.5 88.7 62,667 83.9 92.3 1126 98.8 98.8 38,379 82.6 92.2 21,502 80.9 91.2
4 1159 831,190 90.7 90.0 19,185 87.4 88.1 10,508 94.5 94.5 19,131 84.9 86.6 8416 83.0 84.5
5 147 1,526,000 90.6 90.2 17,017 86.1 88.3 53,860 96.3 96.3 466 83.4 87.4 61,587 82.0 86.4
Total 4326 2,878,665 86.4 89.7 157,452 84.6 93.0 63,575 96.3 96.3 121,118 81.7 92.3 27,217 82.6 91.5
a
All cost and savings gures are in euros.
1
6
E
.
P
o
r
r
a
s
,
R
.
D
e
k
k
e
r
/
E
u
r
o
p
e
a
n
J
o
u
r
n
a
l
o
f
O
p
e
r
a
t
i
o
n
a
l
R
e
s
e
a
r
c
h
x
x
x
(
2
0
0
7
)
x
x
x
x
x
x
A
R
T
I
C
L
E
I
N
P
R
E
S
S
P
l
e
a
s
e
c
i
t
e
t
h
i
s
a
r
t
i
c
l
e
i
n
p
r
e
s
s
a
s
:
P
o
r
r
a
s
,
E
.
,
D
e
k
k
e
r
,
R
.
,
A
n
i
n
v
e
n
t
o
r
y
c
o
n
t
r
o
l
s
y
s
t
e
m
f
o
r
s
p
a
r
e
p
a
r
t
s
a
t
a
r
e
n
e
r
y
:
.
.
.
,
E
u
r
.
J
.
O
p
e
r
.
R
e
s
.
(
2
0
0
7
)
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
j
o
r
.
2
0
0
6
.
1
1
.
0
0
8
r ll rate realized ll rate, same observations as for
rCSL
Savings (
*
) vs. (C) Dierence between the holding
and ordering costs of the current system
with respect to (N), (P), (E) and (W)
5.1. Optimization results for the system as a whole
From the results at the bottom of Table 4 we can
see that for the aggregated values across all classes
considered in the system, all the models outper-
formed the current system under the ex-post
approach, achieving lower total costs and higher ll
rates, with overall CSL within 5 1% of the current
CSL. For the dierent models under the ex-post
approach, we make the following observations:
The normal model achieved the best overall per-
formance among the models, achieving the highest
total savings, although for demand class 1 the
empirical and Willemain models performed very
similar with respect to the normal. For this class
the Poisson model performed the worst of all mod-
els. Total savings for the normal model were of 8.4%
over a total current cost of 15.5 million euros for the
5-year period, and with overall ll rate of 92.6% ver-
sus a current ll rate of 89.2%. The cycle service
level achieved by this model was 85.1%, slightly
lower than the one of the current system.
The empirical model outperformed Willemains
model in terms of total savings (1.05 vs. 0.96 million
euros), both achieving overall ll rates above 90%
and similar overall cycle service levels. To explain
the dierence in performance of these models, we
found that the Willemain model evaluates slightly
higher re-order points to achieve the same service
level as the ones of the empirical model. This
because the Willemain method produces a larger
range of dierent lead time demand values, whereas
the empirical method uses the exact values observed
in the data set to construct the LTD distribution.
Therefore, larger deviations are expected in the
LTD distribution based on the Willemain model
and expensive individual items are likely to have a
higher eect with respect to holding costs. In this
respect, observe that the dierence in total savings
between the two models comes from few classes
with expensive items (e.g. classes 135, 235, 435),
whereas for other classes both models performed
very similar or the Willemain model performed bet-
ter (see Table 11 in the appendix). On the other
hand the normal model uses a continuous distribu-
tion of LTD, which is smoother than the empirical
or the Willemains one, making it less expensive to
match or improve the target service levels.
As for the ex-ante approach, we have the follow-
ing conclusions (see bottom of Tables 5 and 12 of
the appendix):
As the ex-ante approach uses a reduced data set
for tting purposes, the optimization of the system
became more dicult than when using the ex-post
approach, which could better match or improve
the target service levels at lower total costs. Recall
that fewer items are used under this approach since
we only consider items with at least two positive
demands during the tting period. Despite these
limitations, the results obtained from ex-ante
approach conrmed the general conclusions derived
under the ex-post approach. First of all observe that
the current system was outperformed by all models
although with lower savings.
Second, under this approach again the normal
model achieved the best performance among all
models, with total savings of 5.5% with respect to
the current system and overall ll rate of 93% (over
a current one of 89.7%).
Next, as in the ex-post approach, the empirical
model achieved higher savings than the Willemain
model, both with realized ll rates above 90%. We
can see that again the dierence comes from the very
same classes as the ones discussed above for the ex-
post approach.
Finally, for demand class 1, the Poisson model
performed the worst of all models, which is the same
result found under the ex-post approach.
5.2. Optimization results by demand, criticality and
price classes
From the optimization results aggregated by
demand, criticality and price classes presented in
Tables 4 and 5, we can see the eect of the dierent
methods on particular classes. As we had more
information available for the individual classes
under the ex-post approach, we give the following
observations under this approach, unless otherwise
indicated.
All the models outperformed the current system
for the dierent aggregated classes with positive sav-
ings and higher ll rates, except for demand class 3,
where the empirical and the Willemain models had
higher total costs and lower ll rates; criticality class
1, where the empirical model yielded higher total
costs; and price class 2, where the Poisson model
ARTICLE IN PRESS
achieved also higher total costs. For the latter two,
however, the models achieved the same or higher ll
rates with respect to the current system within 1%
precision. In the ex-ante approach there are few
more classes where the models could not outper-
form the current system (demand class 4, criticality
class 3 and price class 5, with varied results for the
dierent models).
From the optimization results by demand classes,
we can see that, according to what one would
expect, for all models and for the current system
the achieved service levels (CSLs and ll rates)
decreased as demand for the items increased. For
low demand items (demand classes 1 and 4), all
models achieved high CSLs and ll rates, with sav-
ings ranging from 6.2% to 20% with respect to the
current system. However for high demand items
(demand classes 2 and 3) the normal model achieved
much better results than the other models, with total
combined savings of 4.3%, as compared to the cor-
responding savings of 1.8% and 1.5% of the empir-
ical and Willemain models.
From the results aggregated by criticality classes
we conclude that for low to medium criticality all
models performed well. However for high critical
items the normal model performed much better
than the others. For high critical items under the
ex-ante approach, observe that the current system
was outperformed by the models but at the expense
of higher costs. For example the normal model
achieved a ll rate 5% higher than the one of the
current system but with 0.35% higher costs. (For a
more specic examination of the relation between
criticality and service levels see analysis below.)
Finally, from the results by price classes, we can
see that the models performed similarly, except for
price classes 2 and 3 where the Poisson model
achieved negative or low total savings.
5.3. A closer look at a class with high dierences:
explaining the results
We want to give in this section some specic
results for one class where we observed large dier-
ences between the models. By doing this we want to
explain the dierences in the general results dis-
cussed in the previous sections. Accordingly, we
chose class 215 as an illustrative example (see opti-
mization results for this class in Table 11 of the
Appendix). Notice that for this class the normal
model was the only one achieving positive savings.
The empirical model yielded lower total costs than
the Willemain model, though it achieved a lower ll
rate than the current system.
In Table 6, we present the dierent steps in the
optimization of four typical items inside class 215,
with price per item of 20,000 euros. These items
are representative of the 15 most expensive items
in this class, with similar lead times and total
demands, where the total number of items was 39.
In Table 6 the following additional results are pre-
sented for each item and model, along with the real-
ized CSL, the realized ll rate and the total costs:
(s, S) the (min, max) parameters of the current
system
s the re-order point corresponding to a target
ll rate (t ll rate)
#c total number of inventory cycles completed
by the item
#soc number of stock out cycles for the given re-
order point
#D total number of units demanded
#S total number of units supplied for the given
re-order point.
After running an optimization procedure for
these items, we obtained the results shown at the
bottom of Table 6. As we can see from these results,
the normal model is the only one for which a target
ll rate can be identied for which the combined
performance of all the items produced an optimized
result with respect to the current system (a target ll
rate of 76%). For the empirical model, since we had
fewer available re-order points, this match is more
dicult to obtain for a particular target ll rate.
On the other hand, the Willemain model produces
a greater range of re-order points as compared to
the normal or the empirical model. However the tar-
get ll rate values for the dierent re-order points do
not allow an optimal combination for a specic
optimal target ll rate.
The previous analysis explains the high variabil-
ity that we observe in the optimized results for some
of the classes presented in Tables 4 and 5. Although
this is a drawback of the optimization procedure
due to the decomposition approach, it allows a trac-
table analysis of large scale systems like the one
related to the present study. Since the method uses
the same target parameters for group of items, the
stock control is less cumbersome than in the case
of parameters set at an individual level. An alterna-
tive approach is discussed in Section 5.8, where we
optimize class 215 using an item level optimization.
ARTICLE IN PRESS
Table 6
Optimization results for some items in class 215
Item # Current system (C) Normal model (N) Empirical model (E) Willemains model (W)
(s, S) #c #soc #D #S Total
costs
CSL
(%)
ll
rate
(%)
t ll
rate
(%)
s #c #soc #D #S Tot.
costs
rCSL
(%)
r ll
rate
(%)
t ll
rate
(%)
costs
rCSL
(%)
r ll
rate
(%)
t ll
rate
(%)
costs
rCSL
(%)
r ll
rate
(%)
870
(L = 1)
(1, 2) 2 0 3 3 44,655 100 100 1 0 2 1 3 2 22,155 50 67 1 0 2 1 3 2 22,155 50 67 1 0 2 1 3 2 22,155 50 67
86 1 2 0 3 3 44,655 100 100 95 1 2 0 3 3 44,655 100 100 90 1 2 0 3 3 44,655 100 100
99 2 2 0 3 3 67,572 100 100 94 2 2 0 3 3 67,572 100 100
97 3 2 0 3 3 90,489 100 100
99 4 2 0 3 3 113,410 100 100
100 6 2 0 3 3 159,240 100 100
11216
(L = 6)
(2, 2) 2 0 4 4 44,239 100 100 1 0 2 2 4 2 26,322 0 50 1 0 2 2 4 2 26,322 0 50 1 0 2 2 4 2 26,322 0 50
43 1 2 0 4 4 44,239 100 100 53 2 2 0 4 4 67,155 100 100 31 1 2 0 4 4 44,239 100 100
83 2 2 0 4 4 67,155 100 100 52 2 2 0 4 4 67,155 100 100
97 3 2 0 4 4 90,072 100 100 71 3 2 0 4 4 90,072 100 100
100 5 2 0 4 4 135,910 100 100 85 4 2 0 4 4 112,990 100 100
93 5 2 0 4 4 135,910 100 100
97 6 2 0 4 4 158,820 100 100
99 7 2 0 4 4 181,740 100 100
100 12 2 0 4 4 296,320 100 100
11186
(L = 3)
(2, 2) 2 0 4 4 49,239 100 100 1 0 2 2 4 2 28,822 0 50 1 0 2 2 4 2 28,822 0 50 1 0 2 2 4 2 28,822 0 50
62 1 2 0 4 4 49,239 100 100 78 2 2 0 4 4 72,155 100 100 65 1 2 0 4 4 49,239 100 100
95 2 2 0 4 4 72,155 100 100 76 2 2 0 4 4 72,155 100 100
100 3 2 0 4 4 95,072 100 100 87 3 2 0 4 4 95,072 100 100
95 4 2 0 4 4 117,990 100 100
99 5 2 0 4 4 140,910 100 100
100 7 2 0 4 4 186,740 100 100
5982
(L = 8)
(2, 2) 2 1 4 2 40,072 50 50 1 0 2 1 4 1 23,405 50 25 1 0 2 1 4 1 23,405 50 25 1 0 2 1 4 1 23,405 50 25
32 1 2 1 4 2 40,072 50 50 34 2 2 1 4 3 62,572 50 75 14 1 2 1 4 2 40,072 50 50
76 2 2 1 4 3 62,572 50 75 96 4 2 0 4 4 107,990 100 100 40 2 2 1 4 3 62,572 50 75
94 3 2 0 4 4 85,072 100 100 63 3 2 0 4 4 85,072 100 100
99 4 2 0 4 4 107,990 100 100 80 4 2 0 4 4 107,990 100 100
100 5 2 0 4 4 130,910 100 100 91 5 2 0 4 4 130,910 100 100
96 6 2 0 4 4 153,820 100 100
99 7 2 0 4 4 176,740 100 100
100 10 2 0 4 4 245,490 100 100
Opt. results 8 1 15 13 178,205 87.5 86.7 76 8 2 15 13 178,205 75.0 86.7 95 8 1 15 14 246,537 87.5 93.3 70 8 1 15 14 223,621 87.5 93.3
E
.
P
o
r
r
a
s
,
R
.
D
e
k
k
e
r
/
E
u
r
o
p
e
a
n
J
o
u
r
n
a
l
o
f
O
p
e
r
a
t
i
o
n
a
l
R
e
s
e
a
r
c
h
x
x
x
(
2
0
0
7
)
x
x
x
x
x
x
1
9
A
R
T
I
C
L
E
I
N
P
R
E
S
S
P
l
e
a
s
e
c
i
t
e
t
h
i
s
a
r
t
i
c
l
e
i
n
p
r
e
s
s
a
s
:
P
o
r
r
a
s
,
E
.
,
D
e
k
k
e
r
,
R
.
,
A
n
i
n
v
e
n
t
o
r
y
c
o
n
t
r
o
l
s
y
s
t
e
m
f
o
r
s
p
a
r
e
p
a
r
t
s
a
t
a
r
e
n
e
r
y
:
.
.
.
,
E
u
r
.
J
.
O
p
e
r
.
R
e
s
.
(
2
0
0
7
)
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
j
o
r
.
2
0
0
6
.
1
1
.
0
0
8
5.4. Service level vs. class structure
Although service levels would usually be reected
by dierent price or criticality levels, we could not
observe for the current system (ex-post results) a
consistent pattern in this respect. In order to have
a better picture of this result, we present in Fig. 5
the aggregated current service levels in a two-entry
matrix for price and criticality. We would expect
that as price increases for the same criticality level
the corresponding service levels would decrease.
On the other hand, for the same price, we expect
that as criticality increases the service level also
increases. However we observe that often expensive
items have higher associated service levels than
cheap ones, and low critical items have higher ser-
vice levels than high critical items. For some other
cases we could observe a more consistent behavior
within the line of expectation, but a general conclu-
sion in this respect is far from obvious. We conclude
that the inconsistencies observed are maybe the
result of a wrong judgment on the part of the stock
analyst when deciding on the re-order points for the
items. Also rounding up the lead times could have
an eect, since by doing this we are generally evalu-
ating worst service levels than the ones experienced
by the system in reality.
5.5. Optimization curves: Cost vs. service level
In Fig. 6, we show a couple of typical optimiza-
tion curves for a particular class. The upper plot
corresponds to the exchange curve between target
and realized ll rates. The current policy is repre-
sented by a horizontal line. The lower picture shows
the total holding and ordering costs as a function of
91.1% 93.3% 89.3% 93.2%
90.4% 89.6% 91.4% 94.4%
96.6% 91.4% 87.8% 95.8%
Current fill rates
91.7% 93.4% 92.2% 86.6%
91.8% 89.3% 89.8% 91.4%
96.2% 91.7% 90.1% 95.0%
Current CSLs
p>2112 169<p<2112 13.6<p<169 p<13.6
91.1% 93.3% 89.3% 93.2% High
90.4% 89.6% 91.4% 94.4% Medium
96.6% 91.4% 87.8% 95.8% Low
Current fill rates
91.7% 93.4% 92.2% 86.6% High
91.8% 89.3% 89.8% 91.4% Medium
96.2% 91.7% 90.1% 95.0% Low
Current CSLs
p>2112 169<p<2112 13.6<p<169 p<13.6 Criticality
Price in
euros
Criticality
increases
Price increases
Fig. 5. Double-entry matrices for CSLs and ll rates of the
current system.
0 10 20 30 40 50 60 70 80 90 100
60
70
80
90
100
Realized fill rate vs. Taget fill rate
Target fill rate
R
e
a
l
i
z
e
d

f
i
l
l

r
a
t
e
60 65 70 75 80 85 90 95 100
0
2
4
6
8
10
x 10
5
Total costs vs. Realized fill rate
Realized fill rate
T
o
t
a
l

C
o
s
t
s
Current policy
Current policy
(E)
(E)
(N)
(N)
(W)
(W)
Fig. 6. Optimization curves for class 233 (ex-post analysis).
ARTICLE IN PRESS
the realized ll rate for the current system and for
the normal, empirical and Willemain models. Obvi-
ously the current policy is represented by a single
point. The plots correspond to class 233 (733 items
with total individual demands larger than 1 and
total demand lower than 60, medium criticality
and 13.6 < price 6 169 euros), and the precision of
the graphs is 1%.
As we can see from the lower plot, the point asso-
ciated to the current system lies above the curves for
the three models. Therefore we can achieve higher
ll rates with positive savings for the three models
(see numerical results for this class in Table 11 of
the Appendix). Notice that the normal curve is less
steep than the ones for the empirical and Willemain
models in the optimization region (around a real-
ized ll rate of 90%), therefore achieving higher sav-
ings for this class. This behavior of the normal
model illustrates the better performance of this
model over the others for some relevant classes, as
we observed the same eect discussed here. Observe
also that for this class the Willemain curve is
above the empirical one, which illustrates the bet-
ter performance of the latter with respect to the
former. Finally, notice that the normal model
needed a higher target ll rate to achieve its optimal
realized ll rate than the other models. As before,
we observed the same eects in other relevant clas-
ses. From this we can conclude that the empirical
and Willemain models are more accurate in terms
of predictability of the service levels. Although in
this case the three models yielded a good perfor-
mance over the current policy, for other classes the
optimization was more dicult, as we showed in
Section 5.3.
In Fig. 7, we present the corresponding optimiza-
tion curves for class 135 (demand class 1, medium
criticality and price >169 euros). Here we can com-
pare the performance of the Poisson model with
respect to other models. As we can see from the
lower plot the performance curves for the four mod-
els are very similar around the optimization region
(realized ll rate of 9597%), all of them clearly out-
performing the current policy. From the upper plot
we can see again that the predictability of the nor-
mal model is the worst of all, achieving higher ll
rates for lower target values than the other models.
Notice that for high ll rates the Poisson model is
the most accurate in this respect. Observe however
that in terms of the optimization performance this
does not have a great impact for the normal or
the other models.
93 94 95 96 97 98 99 100
1
0
2
3
4
x 10
6
Current policy
(E)
(P)
(N)
(W)
Current policy
(E)
(N)
(W)
(P)
0 10 20 30 40 50 60 70 80 90 100
92
90
94
96
98
100
Target fill rate
R
e
a
l
i
z
e
d

f
i
l
l

r
a
t
e
Total costs vs. Realized fill rate
Realized fill rate
T
o
t
a
l

C
o
s
t
s
Fig. 7. Optimization curves for class 135 (ex-post analysis).
ARTICLE IN PRESS
5.6. Accuracy of the models and variability within the
classes
As the number of inventory cycles completed by
each item inside a given class may vary greatly, we
expect variations inside the classes between the tar-
get and the realized ll rates for the individual items.
In order to give an estimation of the accuracy of the
models, in Table 7 we present the average and max-
imum values of the ll rate MADs, aggregated for
the dierent demand classes. For the calculation
the current ll rates were used as target values under
the ex-post approach. As we can see from the
results, the MAD calculation associated with high
demand classes yielded higher average and maxi-
mum values. From the average results we can con-
clude that the accuracy of the models varies
between 4% and 17% as given by the MADs, which
is due to the low number of demand occurrences.
Another measure of variability is the one pro-
vided by the weighted variances of the realized ser-
vice levels, as dened in Section 4.3. This measure is
more related to the performance of the system than
the MAD calculation, as it assesses the variability of
the realized service levels for a given set of parame-
ters within a class. In Table 8, we give the values
of the variances aggregated for the demand classes
for the dierent models. As we can see from the
results the service level variances associated with
high demand classes are higher than for low demand
classes. In general, for the same demand class the
variances of realized CSLs and ll rates are within
similar ranges for the dierent models.
5.7. Individual items
Although the optimization of the system has so
far been performed at class level, where target ser-
vice levels are the same for all items inside the class,
it is interesting to look at the dierence between the
models for individual items. With this idea we show
in Fig. 8 a number of interesting plots for two typ-
ical items belonging to classes 334 and 124. These
plots are analogous to the ones shown in Figs. 5
and 6 but at an item level. The data for these items
is given in Table 9. As we can see from the plots the
curves of item #1835 are smoother than the ones of
item #7559. Since demands for the latter are of unit
size, the steps in realized service levels and associ-
ated costs are bigger for a given target service level.
For some classes these discrete jumps in realized
service levels for the items made the optimization
Table 7
Average and maximum MADs for target ll rates
Demand
class
Normal model (N) Poisson model (P) Empirical model (E) Willemain model (W)
Average
(%)
Maximum
(%)
Average
(%)
Maximum
(%)
Average
(%)
Maximum
(%)
Average
(%)
Maximum
(%)
1 5.5 15.2 5.8 15.4 5.5 15.2 5.8 15.2
2 12.3 18.2 13.2 20.9 14.0 23.0
3 12.9 24.6 17.6 36.3 15.3 24.6
4 4.3 8.8 4.4 8.5 4.4 8.5
Table 8
Variances of realized ll rates and CSLs
Demand
class
Normal model (N) Poisson model (P) Empirical model (E) Willemain model (W)
Average
(%)
Maximum
(%)
Average
(%)
Maximum
(%)
Average
(%)
Maximum
(%)
Average
(%)
Maximum
(%)
Variance of CSLs
1 1.2 3.8 0.9 3.2 1.2 3.8 1.3 3.8
2 6.9 20.2 9.0 17.8 9.9 20.7 9.9 20.7
3 5.3 12.8 9.1 14.5 8.0 11.3
4 1.0 3.0 1.0 2.5 0.9 2.5
Variance of ll rates
1 1.1 4.2 0.9 3.2 1.0 4.2 1.2 4.2
2 4.6 10.5 2.2 6.2 4.6 9.7
3 8.8 35.2 6.4 24.3 4.2 18.2
4 0.6 3.2 0.6 3.2 0.6 3.2
ARTICLE IN PRESS
very dicult, as it was the case for class 124. As for
item #1835 the normal and Willemain models
yielded very smooth curves in spite of the few
observed demands. For the class where this item
belongs this proved to be advantageous for these
two models with respect to the empirical model.
Note that there seems to be some regularity in the
demands, which could be due to preventive mainte-
nance, but the data do not provide information
about it.
5.8. Optimization using a marginal approach (item
level optimization)
As we saw in the analysis presented in Section
5.3, the optimization method based on the decom-
position approach may not nd an optimal solution
for classes where items have large dierences in
holding costs. This because we use a single target ll
rate for all items inside the class. As this eect
depends on the ratio of most expensive items to
the cheapest ones, we try to minimize it by cluster-
ing the items in the price classes dened for the
study. However, as we saw in the optimization
results, for some classes we could not outperform
the current policy using the proposed methodology.
In order to overcome this problem, one should opti-
mize the dierent classes using a marginal approach.
To this end, the optimal re-order point for each item
should be found independently using a measure of
the contribution of the item to the overall ll rate
of the class per monetary unit invested. As the pro-
cedure considers each item individually, for large
scale systems the associated computation eort
may be large and therefore the implementation of
such a method is not feasible. Hence, for these cases
an approach of the type presented in this study may
be more suitable.
(N)
(E)
(W)
Item #1835, Class 334
0 50 100
0
50
100
target fill rate
0 50 100
target fill rate
r

f
i
l
l

r
a
t
e
r

f
i
l
l

r
a
t
e
r fill rate
r fill rate
r fill rate
0 50 100
0
50
100
target fill rate
0 50 100
0 50 100 0 50 100
target fill rate
r

f
i
l
l

r
a
t
e
0
50
100
r

f
i
l
l

r
a
t
e
20 40 60 80 100
0
1
2
x 10
4
10
4
10
4
x
x
Tot. Costs vs. r fill rate
T
o
t
a
l

c
o
s
t
s
T
o
t
a
l

c
o
s
t
s
T
o
t
a
l

c
o
s
t
s
T
o
t
a
l

c
o
s
t
s
T
o
t
a
l

c
o
s
t
s
T
o
t
a
l

c
o
s
t
s
T
o
t
a
l

c
o
s
t
s
0
2
4
0
1
2
3
(N)
(E)
(W)
(P)
90
95
100
0 50 100
target fill rate
r

f
i
l
l

r
a
t
e
90
95
100
0 50 100
target fill rate
r

f
i
l
l

r
a
t
e
90
95
100
0 50 100
target fill rate
r

f
i
l
l

r
a
t
e
90
95
100
90 95 100
0
2000
4000
Total Costs vs. Realized fill rate
r fill rate
90 95 100
r fill rate
90 95 100
r fill rate
90 95 100
r fill rate
0
1000
2000
0
5000
1000
500
1500
Fig. 8. Plots for typical items.
Table 9
Data for items depicted in Fig. 8
L = 31 days (1 month)
Price = 236 euros
Demands 10 24 76
Period 12 49 52
L = 44 days (2 months)
Price = 485 euros
Demands 1 1 1 1 1 1 1 1 1 1 1
Period 6 9 12 13 18 26 29 39 47 49 55
ARTICLE IN PRESS
Accordingly, we present in this section an alter-
native optimization approach for the system based
on the marginal approach described above, using
the ll rate as the optimization criterion. This proce-
dure is similar to the one presented in Muckstadt
(2005), where a Poisson model is used to optimize
the system subject to a budget constraint. First we
introduce the following notation:
D
j
average annual demand of item j
s
j
(b) re-order point of item j associated with a ll
rate b
b
j
(s
j
) expected ll rate of item j corresponding to
a re-order point s
j
as given by the LTD dis-
tribution (see Step 2 of method in Section
4.3)
s
(k)
the kth set of given re-order points for a
group of items, say [s
1
, s
2
, . . . , s
n
]
E(b
p
(s
(k)
)) expected ll rate for class p associated
with the kth set of re-order points for the
items inside the class, given by
Eb
p
s
k

n
j1
b
j
s
k
j
D
j
n
j1
D
j
:
Note: For illustrative purposes, we implemented this
method for some of the classes for which we could
not outperform the current policy using the original
methodology.
5.8.1. Optimization method using expected marginal
contributions to the class ll rate
1. For each class p, say with n items, consider an ini-
tial list s
(0)
of available re-order points as given
by the corresponding LTD distribution of each
item, say for b = 0%:
s
0
s
0
1
; s
0
2
; . . . ; s
0
n
:
2. For each item j evaluate the expected marginal
contribution to the ll rate of the class per incre-
mental euro invested if sj
k
is increased to s
k1
j
,
where s
k1
j
is the next available re-order point
given by the LTD distribution of the item (in
the case of the normal model this is equivalent
to increase the re-order point by one unit). That
is, evaluate
Ds
k1
j

D
j
b
j
s
k1
j
b
j
s
k
j

h
j
s
k1
j
s
k
j

;
where h
j
is the annual unit holding cost of item j. T
a
b
l
e
1
0
O
p
t
i
m
i
z
a
t
i
o
n
r
e
s
u
l
t
s
f
o
r
s
e
l
e
c
t
e
d
c
l
a
s
s
e
s
u
s
i
n
g
t
h
e
m
a
r
g
i
n
a
l
a
p
p
r
o
a
c
h
I
t
e
m
c
l
a
s
s
C
u
r
r
e
n
t
s
y
s
t
e
m
(
C
)
N
o
r
m
a
l
m
o
d
e
l
(
N
)
P
o
i
s
s
o
n
m
o
d
e
l
(
P
)
E
m
p
i
r
i
c
a
l
m
o
d
e
l
(
E
)
W
i
l
l
e
m
a
i
n
m
o
d
e
l
(
W
)
T
o
t
.
c
o
s
t
s
S
e
r
v
i
c
e
l
e
v
e
l
(
%
)
M
e
t
h
o
d
S
a
v
i
n
g
s
S
e
r
v
i
c
e
l
e
v
e
l
(
%
)
S
a
v
i
n
g
s
S
e
r
v
i
c
e
l
e
v
e
l
(
%
)
S
a
v
i
n
g
s
S
e
r
v
i
c
e
l
e
v
e
l
(
%
)
S
a
v
i
n
g
s
S
e
r
v
i
c
e
l
e
v
e
l
(
%
)
C
S
L
l
l
r
a
t
e
(
N
)
v
s
.
(
C
)
r
C
S
L
r
l
l
r
a
t
e
(
P
)
v
s
.
(
C
)
r
C
S
L
r
l
l
r
a
t
e
(
E
)
v
s
.
(
C
)
r
C
S
L
r
l
l
r
a
t
e
(
W
)
v
s
.
(
C
)
r
C
S
L
r
l
l
r
a
t
e
1
6
5
3
1
6
,
2
8
0
9
5
.
4
9
5
.
4
M
a
r
g
i
n
a
l
2
3
2
9
5
.
4
9
5
.
4
2
3
2
9
5
.
4
3
9
5
.
4
3
3
7
0
6
9
5
.
4
9
5
.
4
8
0
8
9
6
.
1
9
6
.
1
D
e
c
o
m
p
o
s
i
t
i
o
n
5
9
9
5
.
4
9
5
.
4
5
9
9
5
.
4
0
9
5
.
4
0
5
9
9
5
.
4
9
5
.
4
2
6
8
7
9
6
.
1
9
6
.
1
2
1
5
8
8
0
,
4
1
0
8
3
.
2
8
1
.
3
M
a
r
g
i
n
a
l
5
2
,
2
2
2
7
8
.
8
8
8
.
8
9
4
3
6
7
2
.
1
8
2
.
1
1
7
0
5
6
2
.
5
8
1
.
3
D
e
c
o
m
p
o
s
i
t
i
o
n
1
9
,
2
3
4
7
8
.
8
8
7
.
9
7
5
5
8
5
3
.
8
7
7
.
2
8
0
2
7
6
9
.
2
8
4
.
8
3
3
3
1
0
1
,
2
0
0
7
4
.
1
7
5
.
6
M
a
r
g
i
n
a
l
1
7
,
8
1
5
6
8
.
2
7
5
.
0
2
9
,
9
3
9
5
8
.
5
7
2
.
9
2
9
,
8
0
5
6
7
.
2
7
4
.
4
D
e
c
o
m
p
o
s
i
t
i
o
n
2
6
5
8
6
0
.
4
7
8
.
6
1
6
7
3
4
4
.
9
6
7
.
6
3
6
9
7
4
5
.
5
7
3
.
6
ARTICLE IN PRESS
3. Increase the re-order point of the item with the
highest value of Ds
k1
j
, say item l, to get a
new list of re-order points, given by
s
k1
s
k
1
; s
k
2
; . . . ; s
k1
l
; . . . ; s
k
n
:
4. Simulate the system with the list of re-order
points given by s
(k+1)
and evaluate the realized
CSL, the realized ll rate, the total costs and
the total savings for the class.
5. Compare with the current policy as in step 4 of
the previous optimization method (Section 4.3).
If the realized service levels for the class are
below the current ones, repeat steps (2)(4) until
the service levels are within acceptable levels or
until the total costs for the class exceed the cur-
rent ones.
In Table 10, we present the results for the optimi-
zation of classes 165, 415 and 533 using the above
method. We also include the results of the optimiza-
tion using the decomposition approach of Section
4.3. As we can see from the results, for all three clas-
ses the models yielded better results with respect to
the previous optimization method, except in the
case of the empirical model for classes 165 and
415, which achieved higher total costs than before.
Notice that as we are looking at the optimal individ-
ual re-order points, the new method can take better
advantage of the LTD distributions modeled
according to the dierent methods. However, as
for the previous optimization method, some prob-
lems may arise as we are limited by the range of
re-order point values associated with the LTD dis-
tribution. This eect is more clearly manifested in
the case of the empirical model, as we observed in
the results. Recall that the LTD distribution in the
case of the empirical model is determined using only
available demand values from the demand data, and
thus the range of re-order points is lower than for
Willemains or the Normal models.
6. Conclusions
In this study we present a methodology for
the empirical test of dierent inventory models
with actual data for spare parts, using two dier-
ent approaches, namely an ex-post and an ex-ante
approach. Although in practice the ex-ante
approach is of more relevance than the ex-post,
as information for spare parts is always limited it
is also more dicult to implement. As a result,
for the ex-post approach we showed that the cur-
rent stock control practices can be improved by
using the theoretical models considered, however
using the ex-ante approach the results were less
conclusive. Consequently, total savings of up to
6.4% in holding costs can be achieved by having
a better inventory control. We have to note how-
ever, that using rounded-up values for the lead
times generally hurts the realized service levels. A
more rened study would require the use of more
accurate values for the lead times. Of course this
would also require more specic information about
the demand data, which is often in practice dicult
to obtain.
According to our study, we can derive the follow-
ing general conclusions on the inventory control of
spare parts:
1. We observed erratic high demands partially
caused by preventive maintenance, which needs
to be identied for a better stock control of the
parts.
2. Spare parts present low to very low demand
values (in some cases no demand observations
were observed during 5 years). With so limited
data the application of theoretical models is
troublesome.
3. Predictability of theoretical models is an impor-
tant issue, as it measures the expected service
level of a policy with given parameters when con-
fronted with real data. We observed large dier-
ences between target and realized service levels,
with maximal values of up to 40%.
4. ERP packages as SAP R/3 do not provide the
right tools for the control of spare parts, since
they only include the cycle service level (CSL)
as service measure. In this case, a more appropri-
ate measure is provided by the ll rate.
5. A theoretical analysis of the type presented in
this study may be hampered by non-stationarity
of the demand data, as lead time demand pat-
terns may change over time.
6. The information provided on the demand sizes is
an important issue, as they are often not related
to the number of items installed. As a conse-
quence of this, compound-Poisson models are
dicult to use.
7. Inventory models can save money and improve
service levels, even in the ex-ante approach.
For the models under study we have the follow-
ing general conclusions:
ARTICLE IN PRESS
Table 11
Optimization results for the system using the ex-post approach
Item
class
# of
items
Tot. costs
(euros)
Service level
(%)
Savings Service level
(%)
(%)
(%)
(%)
tCSL t ll
rate
(N) vs.
(C)
rCSL r ll
rate
(P) vs.
(C)
rCSL r ll
rate
(E) vs.
(C)
rCSL r ll
rate
(W) vs.
(C)
rCSL r ll
rate
113 47 9194 94.3 94.4 1683 93.6 96.5 157 95.8 95.8 1683 93.6 96.5 1683 93.6 96.5
114 60 50,336 87.6 87.8 2671 89.3 90.0 2498 89.4 89.4 2671 89.3 90.0 2671 89.3 90.0
115 23 278,720 92.0 92.0 886 96.0 96.0 886 96.0 96.0 886 96.0 96.0 19,924 94.0 94.0
122 17 2712 100.0 100.0 2101 100.0 100.0 122 100.0 100.0 2101 100.0 100.0 2101 100.0 100.0
123 43 12,908 100.0 100.0 3045 100.0 100.0 2175 100.0 100.0 3667 100.0 100.0 2711 100.0 100.0
124 36 44,260 98.9 98.9 313 99.4 99.4 838 99.4 99.4 313 99.4 99.4 551 99.4 99.4
132 109 11,882 99.6 99.7 9341 100.0 100.0 1857 97.8 97.8 9341 100.0 100.0 9341 100.0 100.0
133 356 93,357 96.6 96.8 26,254 96.2 97.9 11,838 97.9 97.9 29,178 95.9 97.9 27,010 96.0 97.9
134 461 413,830 94.8 94.8 10,447 96.2 96.3 5392 97.3 97.3 13,046 96.9 96.9 14,412 95.7 95.8
135 77 1,821,500 94.1 94.2 633,560 95.9 95.9 633,560 95.9 95.9 613,930 95.9 95.9 599,000 95.9 95.9
142 66 7245 100.0 100.0 5307 100.0 100.0 2667 98.5 98.5 5307 100.0 100.0 5307 100.0 100.0
143 165 44,549 97.5 97.8 13,592 97.5 98.9 4016 99.4 99.4 14,796 97.5 98.9 13,326 97.5 98.9
144 159 141,200 93.6 92.3 21,488 93.8 93.5 20,364 94.5 94.0 21,643 94.0 93.8 17,418 93.3 93.5
145 9 35,389 96.2 96.2 1294 96.2 96.2 1294 96.2 96.2 4027 100.0 100.0 4228 96.2 96.2
152 17 1622 100.0 100.0 1136 100.0 100.0 245 100.0 100.0 1136 100.0 100.0 1136 100.0 100.0
153 39 10,519 94.2 94.8 4108 90.9 96.1 1735 96.1 96.1 4108 90.9 96.1 4108 90.9 96.1
154 76 69,500 93.1 93.1 4021 94.8 94.8 4021 94.8 94.8 4968 94.4 94.4 795 94.0 94.0
155 9 77,338 95.5 95.5 4452 100.0 100.0 4452 100.0 100.0 4452 100.0 100.0 1532 95.5 95.5
162 49 5880 100.0 100.0 4442 100.0 100.0 524 95.9 95.9 4442 100.0 100.0 4442 100.0 100.0
163 157 40,285 98.1 98.1 12,549 98.5 99.2 5705 98.8 98.8 11,052 98.9 99.4 11,659 98.5 99.2
164 204 204,900 92.1 92.0 23,708 93.4 93.3 22,956 93.3 93.2 27,226 93.1 93.0 23,452 93.2 93.2
165 47 316,280 95.4 95.4 59 95.4 95.4 59 95.4 95.4 59 95.4 95.4 2687 96.1 96.1
212 49 6942 84.8 84.4 3018 89.3 99.0 3043 82.1 98.5 2896 85.7 98.6
213 120 47,285 88.8 86.2 1028 85.7 89.7 201 61.5 82.6 1118 69.4 84.3
214 85 173,060 86.5 85.9 3296 80.4 86.3 416 70.2 80.1 340 70.2 80.1
215 39 880,410 83.2 81.3 19,234 78.8 87.9 7558 53.8 77.2 8027 69.2 84.8
222 112 19,791 95.3 98.2 11,836 96.9 99.7 12,777 93.8 99.9 12,778 90.6 99.3
223 129 63,598 94.8 93.5 15,038 91.6 95.9 12,245 91.0 97.4 7650 91.0 96.8
224 60 90,761 92.3 92.0 4808 93.8 93.0 636 89.0 93.2 2696 84.5 88.2
232 411 61,073 88.6 89.2 34,385 84.7 98.3 31,649 91.5 99.6 30,899 92.0 99.3
233 730 366,720 87.9 87.6 69,228 83.4 89.0 19,864 85.3 93.8 9795 83.7 90.9
234 519 1,034,500 87.0 83.9 40,986 85.0 85.2 7146 73.8 83.8 6714 74.4 81.0
235 119 4,200,600 84.1 75.9 80,091 78.7 83.5 13,092 65.0 79.7 25,708 65.4 77.3
242 284 39,537 93.4 94.8 24,393 88.9 99.2 24,153 88.9 99.5 23,351 93.3 99.6
243 388 164,630 90.9 90.2 35,732 86.5 91.9 21,898 87.1 94.6 19,096 86.2 91.5
244 199 261,640 87.2 87.6 12,656 87.4 89.2 929 83.2 88.5 2833 79.7 85.0
2
6
E
.
P
o
r
r
a
s
,
R
.
D
e
k
k
e
r
/
E
u
r
o
p
e
a
n
J
o
u
r
n
a
l
o
f
O
p
e
r
a
t
i
o
n
a
l
R
e
s
e
a
r
c
h
x
x
x
(
2
0
0
7
)
x
x
x
x
x
x
A
R
T
I
C
L
E
I
N
P
R
E
S
S
P
l
e
a
s
e
c
i
t
e
t
h
i
s
a
r
t
i
c
l
e
i
n
p
r
e
s
s
a
s
:
P
o
r
r
a
s
,
E
.
,
D
e
k
k
e
r
,
R
.
,
A
n
i
n
v
e
n
t
o
r
y
c
o
n
t
r
o
l
s
y
s
t
e
m
f
o
r
s
p
a
r
e
p
a
r
t
s
a
t
a
r
e
n
e
r
y
:
.
.
.
,
E
u
r
.
J
.
O
p
e
r
.
R
e
s
.
(
2
0
0
7
)
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
j
o
r
.
2
0
0
6
.
1
1
.
0
0
8
252 28 5593 90.0 91.6 2870 92.9 99.6 3194 85.7 97.5 3041 92.9 98.3
253 76 34,290 86.0 80.9 4197 83.2 90.1 2058 84.2 93.3 1198 81.5 92.1
254 80 147,890 86.7 85.7 12,366 86.0 87.8 3589 79.7 88.6 1147 77.7 85.7
255 14 210,480 93.2 94.2 34,719 93.0 95.7 24,706 93.0 95.7 47,314 93.0 95.7
262 590 112,880 68.7 84.2 60,467 64.5 95.9 55,350 73.9 98.2 59,157 66.1 96.3
263 798 310,520 76.4 80.7 51,941 72.5 86.3 23,856 72.0 87.9 11,960 72.6 86.2
264 314 649,760 83.8 83.7 58,369 84.1 84.8 7215 82.8 86.0 7557 79.1 80.1
265 41 587,700 82.4 79.5 3078 75.7 84.6 1471 70.6 82.5 597 65.4 77.1
313 11 83,779 76.4 88.2 3102 71.7 88.4 7057 35.8 71.4 1469 34.0 74.1
322 25 15,743 66.3 90.2 1761 92.7 99.0 5519 63.6 92.1 5203 63.6 91.4
323 6 8629 86.4 54.9 333 68.0 59.8 7221 28.0 82.8 5964 28.0 80.7
332 144 82,834 80.2 89.7 14,346 86.5 97.8 15,984 78.8 97.2 15,239 80.3 97.5
333 43 101,200 74.1 75.6 2658 60.4 78.6 1673 44.9 67.6 3697 45.5 73.6
334 9 56,697 69.4 75.0 6403 69.0 76.6 620 71.4 77.1 3462 66.7 77.5
342 35 16,182 86.3 93.2 4249 89.1 96.2 4494 89.1 98.2 4482 82.8 97.3
343 9 105,480 75.8 86.4 9147 78.3 85.8 12,775 30.4 76.0 25,708 34.8 70.5
362 287 134,440 77.0 86.9 624 58.9 90.4 11,618 6.4 74.2 8168 6.4 73.7
363 39 80,942 71.4 70.5 20,476 67.6 72.5 443 37.5 73.7 1176 42.6 77.0
413 14 2161 96.9 96.9 73 95.8 96.9 73 95.8 96.9 73 95.8 96.9
414 28 43,355 97.1 97.2 747 98.4 98.6 2347 96.8 97.2 747 98.4 98.6
415 17 129,460 100.0 100.0 312 94.3 95.2 399 97.1 97.6 312 94.3 95.2
423 8 3446 100.0 100.0 1025 100.0 100.0 1203 100.0 100.0 1296 100.0 100.0
424 11 15,405 98.0 98.0 1087 97.7 98.0 723 95.3 96.0 723 95.3 96.0
433 52 14,778 98.1 98.2 95 98.2 98.8 417 98.2 98.8 871 99.1 99.4
434 91 117,410 95.7 95.9 4430 96.3 97.0 2773 96.3 97.0 587 96.3 97.0
435 69 788,730 93.2 93.3 86,946 92.3 94.8 88,500 93.0 95.2 68,706 93.0 95.2
443 36 9741 97.8 98.5 2750 100.0 100.0 2780 100.0 100.0 2250 100.0 100.0
444 39 47,642 97.7 97.8 3602 98.7 98.9 4792 98.7 98.9 4009 98.7 98.9
445 6 32,143 91.7 92.3 6158 100.0 100.0 1316 100.0 100.0 2178 100.0 100.0
454 26 34,420 95.3 95.5 497 94.7 95.5 497 94.7 95.5 497 94.7 95.5
455 7 42,668 100.0 100.0 2534 100.0 100.0 2534 100.0 100.0 2534 100.0 100.0
464 30 40,800 97.5 97.6 1099 96.2 97.6 21 96.2 97.6 1099 96.2 97.6
465 41 353,910 96.6 95.9 0 93.6 95.9 0 93.6 95.9 14,991 96.2 97.2
Total 8494 15,525,060 87.5 89.2 1,307,055 85.1 92.6 703,246 96.8 96.8 1,054,723 82.2 92.7 964,832 81.5 91.4
1 2226 3,693,400 95.6 95.6 774,220 96.2 96.8 703,250 96.8 96.8 762,330 96.3 96.9 751,500 96.0 96.6
2 5185 9,469,700 84.3 86.1 408,750 81.1 90.7 202,860 80.0 91.7 170,000 79.1 89.7
3 608 685,920 77.2 85.7 31,370 70.0 90.2 15,411 37.4 81.4 17,797 37.8 81.7
4 475 1,676,100 96.5 96.6 92,713 96.3 97.3 104,950 96.3 97.3 61,126 96.7 97.5
(continued on next page)
E
.
P
o
r
r
a
s
,
R
.
D
e
k
k
e
r
/
E
u
r
o
p
e
a
n
J
o
u
r
n
a
l
o
f
O
p
e
r
a
t
i
o
n
a
l
R
e
s
e
a
r
c
h
x
x
x
(
2
0
0
7
)
x
x
x
x
x
x
2
7
A
R
T
I
C
L
E
I
N
P
R
E
S
S
P
l
e
a
s
e
c
i
t
e
t
h
i
s
a
r
t
i
c
l
e
i
n
p
r
e
s
s
a
s
:
P
o
r
r
a
s
,
E
.
,
D
e
k
k
e
r
,
R
.
,
A
n
i
n
v
e
n
t
o
r
y
c
o
n
t
r
o
l
s
y
s
t
e
m
f
o
r
s
p
a
r
e
p
a
r
t
s
a
t
a
r
e
n
e
r
y
:
.
.
.
,
E
u
r
.
J
.
O
p
e
r
.
R
e
s
.
(
2
0
0
7
)
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
j
o
r
.
2
0
0
6
.
1
1
.
0
0
8
Table 11 (continued)
Item
class
# of
items
Tot. costs
(euros)
Service level (%) Savings Service level (%) Savings Service level (%) Savings Service level (%) Savings Service level (%)
tCSL t ll
rate
(N) vs.
(C)
rCSL r ll
rate
(P) vs.
(C)
rCSL r ll
rate
(E) vs.
(C)
rCSL r ll
rate
(W) vs.
(C)
rCSL r ll
rate
1 493 1,704,700 88.9 88.3 20,428 87.3 91.8 3541 92.9 92.9 5074 76.1 87.7 16,583 79.5 88.7
2 447 277,250 94.1 95.3 37,882 95.0 97.1 1215 99.8 99.8 28,617 91.2 97.5 21,805 89.8 96.4
3 3190 9,165,100 90.1 89.8 853,480 88.4 92.7 648,930 97.5 97.5 802,150 86.9 93.5 743,200 86.6 92.2
4 1395 905,370 92.6 92.9 107,170 90.7 94.9 20,420 97.2 97.0 83,448 90.0 95.7 54,293 89.9 94.2
5 372 634,320 90.8 89.8 61,002 90.4 93.4 1060 96.1 96.1 30,051 88.6 94.0 54,554 88.0 92.9
6 2597 2,838,300 79.5 85.4 227,090 75.3 90.6 28,077 95.7 95.6 115,530 70.8 89.9 74,396 68.8 88.2
2 2223 524,350 83.2 90.0 179,030 79.4 97.0 5415 97.9 97.9 166,870 75.1 95.7 171,200 73.7 95.0
3 3266 1,608,000 87.9 88.5 245,180 84.7 91.3 25,627 98.2 98.2 115,240 83.3 92.7 73,012 83.3 91.4
4 2487 3,637,300 90.2 89.4 200,410 90.1 90.7 54,393 95.5 95.4 71,791 86.7 90.4 43,397 85.6 88.4
5 518 9,755,400 90.4 88.2 682,440 88.4 91.5 628,640 96.0 96.0 700,820 83.2 89.9 677,220 84.0 89.3
2
8
E
.
P
o
r
r
a
s
,
R
.
D
e
k
k
e
r
/
E
u
r
o
p
e
a
n
J
o
u
r
n
a
l
o
f
O
p
e
r
a
t
i
o
n
a
l
R
e
s
e
a
r
c
h
x
x
x
(
2
0
0
7
)
x
x
x
x
x
x
A
R
T
I
C
L
E
I
N
P
R
E
S
S
P
l
e
a
s
e
c
i
t
e
t
h
i
s
a
r
t
i
c
l
e
i
n
p
r
e
s
s
a
s
:
P
o
r
r
a
s
,
E
.
,
D
e
k
k
e
r
,
R
.
,
A
n
i
n
v
e
n
t
o
r
y
c
o
n
t
r
o
l
s
y
s
t
e
m
f
o
r
s
p
a
r
e
p
a
r
t
s
a
t
a
r
e
n
e
r
y
:
.
.
.
,
E
u
r
.
J
.
O
p
e
r
.
R
e
s
.
(
2
0
0
7
)
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
j
o
r
.
2
0
0
6
.
1
1
.
0
0
8
Table 12
Optimization results for the system using the ex-ante approach
Item
class
# of
items
Tot. costs
(euros)
Service level
(%)
(%)
(%)
(%)
(%)
tCSL t ll
rate
(N) vs.
(C)
rCSL r ll
rate
(P) vs.
(C)
rCSL r ll
rate
(E) vs.
(C)
rCSL r ll
rate
(W) vs.
(C)
rCSL r ll
rate
113 10 1148 95.5 95.5 7 100.0 100.0 49 95.5 95.5 7 100.0 100.0 56 100.0 100.0
114 18 6686 81.1 81.6 513 90.6 92.1 419 81.6 81.6 513 90.6 92.1 555 93.8 94.7
115 1 1030 100.0 100.0 0 100.0 100.0 0 100.0 100.0 0 100.0 100.0 0 100.0 100.0
122 8 675 100.0 100.0 559 100.0 100.0 14 100.0 100.0 559 100.0 100.0 559 100.0 100.0
123 26 3435 100.0 100.0 1938 100.0 100.0 44 100.0 100.0 1938 100.0 100.0 1938 100.0 100.0
124 20 10,834 98.4 98.5 562 100.0 100.0 1555 100.0 100.0 1597 100.0 100.0 167 98.0 98.5
132 38 1800 100.0 100.0 1387 100.0 100.0 682 97.0 97.0 1387 100.0 100.0 1387 100.0 100.0
133 126 17,510 96.7 96.9 6618 95.4 98.4 686 98.0 98.0 6826 94.3 98.0 6239 94.3 98.0
134 160 64,218 94.0 94.0 11,686 94.8 95.1 7638 95.1 95.1 10,819 94.8 95.1 6328 95.9 96.1
135 20 251,940 93.1 93.5 48,588 96.8 96.8 48,588 96.8 96.8 57,403 96.8 96.8 489 93.5 93.5
142 29 1619 100.0 100.0 1161 100.0 100.0 1115 95.9 95.9 1161 100.0 100.0 1161 100.0 100.0
143 73 10,227 97.3 97.7 4001 98.0 99.4 453 100.0 100.0 4086 98.0 99.4 3594 95.9 98.8
144 70 27,042 95.1 95.2 4291 96.3 96.6 686 95.9 95.9 4291 96.3 96.6 2496 96.3 96.6
145 3 4573 83.3 83.3 2405 100.0 100.0 2405 100.0 100.0 2405 100.0 100.0 1334 83.3 83.3
152 3 141 100.0 100.0 83 100.0 100.0 20 100.0 100.0 83 100.0 100.0 83 100.0 100.0
153 18 1993 100.0 100.0 852 100.0 100.0 0 100.0 100.0 852 100.0 100.0 852 100.0 100.0
154 23 9840 97.4 97.4 667 97.2 97.4 448 97.4 97.4 352 97.2 97.4 1707 97.2 97.4
155 1 4221 100.0 100.0 2867 100.0 100.0 2867 100.0 100.0 2867 100.0 100.0 2867 100.0 100.0
162 15 1221 100.0 100.0 943 100.0 100.0 87 85.7 85.7 943 100.0 100.0 943 100.0 100.0
163 39 6267 97.3 97.6 3090 100.0 100.0 985 98.8 98.8 3090 100.0 100.0 3090 100.0 100.0
164 85 40,389 91.5 91.5 3142 92.5 93.2 2869 93.2 93.2 5058 91.8 92.6 4791 91.8 92.6
165 14 37,848 91.4 91.4 4810 94.3 94.3 4810 94.3 94.3 4299 91.4 91.4 4810 94.3 94.3
212 17 1476 80.8 84.3 669 66.7 98.6 761 100.0 100.0 416 100.0 100.0
213 41 8479 85.9 84.1 695 81.4 91.9 349 86.0 93.6 29 69.8 89.4
214 24 19,921 88.4 86.7 559 72.7 78.8 43 84.1 81.4 220 75.0 73.5
215 10 104,870 85.7 74.4 1497 78.6 82.1 15,647 71.4 87.2 9334 64.3 82.1
222 82 6297 97.0 99.2 3800 100.0 100.0 3954 66.7 98.7 4202 100.0 100.0
223 90 20,815 96.0 96.6 1536 97.0 99.1 5053 92.9 97.8 459 94.9 98.0
224 37 24,348 92.6 93.9 2466 93.5 96.4 1760 90.9 93.9 423 87.0 92.1
232 231 16,018 90.7 91.6 8233 82.8 97.7 8330 82.8 97.7 8471 89.7 98.0
233 363 80,282 89.5 88.3 10,956 84.8 91.6 3097 84.6 93.1 26 78.6 90.1
234 247 243,860 86.5 83.1 344 77.4 78.5 353 69.7 75.6 1523 68.1 72.2
235 49 816,360 87.8 90.1 27,678 78.6 82.1 18,160 73.8 78.4 38,595 71.4 77.2
242 198 12,536 93.6 95.5 7838 100.0 100.0 8096 100.0 100.0 8340 100.0 100.0
243 246 51,392 92.8 92.3 11,714 88.1 94.4 8931 88.5 94.4 5584 88.5 93.5
244 124 69,464 89.7 92.2 1919 88.1 89.9 740 85.0 87.1 627 80.8 83.7
(continued on next page)
E
.
P
o
r
r
a
s
,
R
.
D
e
k
k
e
r
/
E
u
r
o
p
e
a
n
J
o
u
r
n
a
l
o
f
O
p
e
r
a
t
i
o
n
a
l
R
e
s
e
a
r
c
h
x
x
x
(
2
0
0
7
)
x
x
x
x
x
x
2
9
A
R
T
I
C
L
E
I
N
P
R
E
S
S
P
l
e
a
s
e
c
i
t
e
t
h
i
s
a
r
t
i
c
l
e
i
n
p
r
e
s
s
a
s
:
P
o
r
r
a
s
,
E
.
,
D
e
k
k
e
r
,
R
.
,
A
n
i
n
v
e
n
t
o
r
y
c
o
n
t
r
o
l
s
y
s
t
e
m
f
o
r
s
p
a
r
e
p
a
r
t
s
a
t
a
r
e
n
e
r
y
:
.
.
.
,
E
u
r
.
J
.
O
p
e
r
.
R
e
s
.
(
2
0
0
7
)
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
j
o
r
.
2
0
0
6
.
1
1
.
0
0
8
Table 12 (continued)
Item
class
# of
items
Tot. costs
(euros)
Service level
(%)
(%)
(%)
(%)
(%)
tCSL t ll
rate
(N) vs.
(C)
rCSL r ll
rate
(P) vs.
(C)
rCSL r ll
rate
(E) vs.
(C)
rCSL r ll
rate
(W) vs.
(C)
rCSL r ll
rate
252 14 655 92.3 95.3 415 100.0 100.0 415 100.0 100.0 415 100.0 100.0
253 40 8238 91.3 89.6 2235 88.9 94.8 841 88.9 96.4 858 88.9 95.3
254 41 27,507 93.8 96.0 553 85.3 89.4 1875 85.3 89.4 940 78.7 86.8
255 2 9770 100.0 100.0 0 100.0 100.0 1765 100.0 100.0 0 100.0 100.0
262 343 34,861 64.3 80.4 20,357 63.6 98.8 19,640 61.4 98.8 19,118 68.2 99.1
263 460 94,819 75.3 80.5 17,881 71.6 87.9 3795 71.4 88.1 759 71.6 87.6
264 198 195,970 86.9 84.9 711 82.8 81.5 1157 79.3 78.0 1627 74.9 73.3
265 16 127,180 79.3 75.0 2150 67.9 71.4 10,862 64.3 73.8 12,382 67.9 76.2
313 7 7632 89.5 97.1 443 73.7 93.6 337 57.9 78.7 63 57.9 82.9
322 24 6481 75.0 92.2 965 94.7 98.4 2333 73.7 95.2 1898 78.9 97.9
323 5 4542 88.9 54.2 1493 75.0 33.4 1436 87.5 35.6 400 87.5 37.3
332 109 19,019 78.8 89.7 1637 84.1 96.8 2970 78.0 95.5 1641 81.7 96.0
333 29 28,546 78.5 92.2 865 64.1 87.0 448 40.6 74.2 239 42.2 76.1
334 6 19,485 80.0 79.6 712 56.3 56.0 683 62.5 62.4 38 62.5 49.7
342 26 5832 86.8 90.7 2369 80.0 91.5 2386 80.0 92.3 1578 80.0 93.6
343 7 9027 84.4 96.6 730 100.0 100.0 2246 40.0 95.4 1670 40.0 95.4
362 235 35,786 72.9 93.5 8166 82.7 96.8 10,127 69.2 95.3 8676 77.9 95.9
363 27 16,405 75.8 67.8 9 48.1 63.1 525 23.1 50.9 64 21.2 51.2
414 4 1258 100.0 100.0 89 100.0 100.0 89 100.0 100.0 89 100.0 100.0
423 5 1055 100.0 100.0 290 100.0 100.0 290 100.0 100.0 203 100.0 100.0
424 4 3432 100.0 100.0 732 77.8 81.8 732 77.8 81.8 732 77.8 81.8
433 22 3387 100.0 100.0 740 100.0 100.0 674 100.0 100.0 609 100.0 100.0
434 42 30,132 98.8 98.9 343 93.8 95.5 310 93.8 95.5 287 93.8 95.5
435 15 108,440 100.0 100.0 8518 92.3 96.0 9666 92.3 96.0 1713 92.3 96.0
443 14 1819 100.0 100.0 681 100.0 100.0 681 100.0 100.0 583 100.0 100.0
444 27 15,882 97.0 97.1 2040 98.0 98.5 2298 98.0 98.5 3000 98.0 98.5
445 1 3025 100.0 100.0 0 100.0 100.0 0 100.0 100.0 0 100.0 100.0
454 14 9630 95.7 95.8 252 94.1 95.8 85 94.1 95.8 632 94.1 95.8
455 3 12,883 100.0 100.0 1152 100.0 100.0 1152 100.0 100.0 1152 100.0 100.0
464 15 11,295 100.0 100.0 1189 100.0 100.0 1189 100.0 100.0 1189 100.0 100.0
465 12 43,899 100.0 100.0 1849 100.0 100.0 1849 100.0 100.0 6570 100.0 100.0
Total 4326 2,878,665 86.4 89.7 157,452 84.6 93.0 63,575 96.3 96.3 121,118 81.7 92.3 27,217 82.6 91.5
3
0
E
.
P
o
r
r
a
s
,
R
.
D
e
k
k
e
r
/
E
u
r
o
p
e
a
n
J
o
u
r
n
a
l
o
f
O
p
e
r
a
t
i
o
n
a
l
R
e
s
e
a
r
c
h
x
x
x
(
2
0
0
7
)
x
x
x
x
x
x
A
R
T
I
C
L
E
I
N
P
R
E
S
S
P
l
e
a
s
e
c
i
t
e
t
h
i
s
a
r
t
i
c
l
e
i
n
p
r
e
s
s
a
s
:
P
o
r
r
a
s
,
E
.
,
D
e
k
k
e
r
,
R
.
,
A
n
i
n
v
e
n
t
o
r
y
c
o
n
t
r
o
l
s
y
s
t
e
m
f
o
r
s
p
a
r
e
p
a
r
t
s
a
t
a
r
e
n
e
r
y
:
.
.
.
,
E
u
r
.
J
.
O
p
e
r
.
R
e
s
.
(
2
0
0
7
)
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
j
o
r
.
2
0
0
6
.
1
1
.
0
0
8
1 800 504,650 95.6 95.7 94,010 96.5 97.2 63,575 96.3 96.3 96,409 96.2 97.1 36,831 96.1 97.2
2 2873 1,975,100 84.8 87.3 53,384 81.3 91.5 10,900 79.1 91.1 18,623 79.6 89.6
3 475 152,750 76.3 90.4 11,873 80.1 92.9 16,379 67.2 90.1 12,119 72.6 90.8
4 178 246,130 98.9 99.0 1816 96.6 97.6 2571 96.6 97.6 3110 96.6 97.6
1 132 152,500 86.5 85.7 540 80.7 90.7 468 87.0 87.0 14,322 87.1 91.5 9752 80.3 88.9
2 301 81,914 94.8 96.5 12,877 97.2 97.8 1496 100.0 100.0 14,668 85.1 96.6 8338 94.6 96.9
3 1457 1,701,000 90.0 90.5 51,385 85.6 91.7 56,230 96.5 96.5 63,253 83.1 91.1 18,236 82.7 89.9
4 818 212,440 93.3 94.3 29,042 93.4 96.1 4659 97.7 97.7 26,538 92.6 95.7 22,704 91.7 94.8
5 159 84,877 94.8 95.2 6130 92.5 95.2 2398 98.7 98.7 2132 92.5 95.6 2947 90.8 94.7
6 1459 645,930 76.3 84.6 58,558 75.6 91.4 8578 94.0 94.0 28,848 71.8 90.5 21,215 74.2 89.9
2 1372 144,420 81.1 90.5 58,581 82.9 98.3 1919 95.2 95.2 63,142 77.6 97.8 58,886 84.3 98.3
3 1648 377,020 87.5 88.7 62,667 83.9 92.3 1126 98.8 98.8 38,379 82.6 92.2 21,502 80.9 91.2
4 1159 831,190 90.7 90.0 19,185 87.4 88.1 10,508 94.5 94.5 19,131 84.9 86.6 8416 83.0 84.5
5 147 1,526,000 90.6 90.2 17,017 86.1 88.3 53,860 96.3 96.3 466 83.4 87.4 61,587 82.0 86.4
E
.
P
o
r
r
a
s
,
R
.
D
e
k
k
e
r
/
E
u
r
o
p
e
a
n
J
o
u
r
n
a
l
o
f
O
p
e
r
a
t
i
o
n
a
l
R
e
s
e
a
r
c
h
x
x
x
(
2
0
0
7
)
x
x
x
x
x
x
3
1
A
R
T
I
C
L
E
I
N
P
R
E
S
S
P
l
e
a
s
e
c
i
t
e
t
h
i
s
a
r
t
i
c
l
e
i
n
p
r
e
s
s
a
s
:
P
o
r
r
a
s
,
E
.
,
D
e
k
k
e
r
,
R
.
,
A
n
i
n
v
e
n
t
o
r
y
c
o
n
t
r
o
l
s
y
s
t
e
m
f
o
r
s
p
a
r
e
p
a
r
t
s
a
t
a
r
e
n
e
r
y
:
.
.
.
,
E
u
r
.
J
.
O
p
e
r
.
R
e
s
.
(
2
0
0
7
)
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
j
o
r
.
2
0
0
6
.
1
1
.
0
0
8
1. Contrary to expectation, the normal model per-
formed very well in the optimization, being the
one that achieved the best overall performance
in both the ex-post and the ex-ante approaches,
achieving total savings of 8.4% (about 1.3 million
euros) over the current system. This in spite of
the fact that the demand process for spare parts
does not generally follow a normal distribution.
We observed that the normal LTD distribution
is less sensitive to changes in service level values,
and thus more conservative re-order points are
advised when using this distribution as com-
pared to Poisson, Willemain or the empirical
LTD distributions. Yet the normal model gives
bad predictions on the actual ll rate to be
achieved.
2. The Poisson model performed equally or worse
than the others for items with 01 demands. We
presume that the Poisson assumption may not
be appropriate, as this model seems to overstock
in several cases. This could be caused by preven-
tive maintenance or by items with wear-out.
3. The empirical model yielded a better overall
performance than the Willemain model, with
total savings of 1.05 and 0.96 million euros,
correspondingly.
4. All models outperformed the current system with
savings and service levels about the same order.
This can be seen as a validation of the implemen-
tation of the dierent models.
5. For the dierent criticality, demand and price
classes considered, all the models performed sim-
ilarly, with only some particular classes where the
normal model yielded higher savings. In this
respect, we did not observe a dierentiated eect
of the use of the dierent models applied to the
various kinds of classes.
All previous conclusions were corroborated by
the ex-ante approach.
Acknowledgement
The authors are thankful to Anne Rijneveld for
doing the case study and providing the data and fur-
thermore to Rutger de Mare and Jeroen Hazeu for
their support in the implementation of the methods
used in this paper.
Appendix
See Tables 11 and 12.
References
Axsater, S., 2000. Inventory Control. Kluwer Academic Pub-
lisher, Boston, MA.
Bookbinder, J.H., Lordahl, A.E., 1989. Estimation of inventory
re-order levels using the bootstrap statistical procedure. IIE
Transactions 21, 302312.
Chopra, S., Meindl, P., 2004. Supply Chain Management:
Strategy, Planning and Operations, second ed. Prentice-Hall
Inc., NJ.
Croston, J.D., 1972. Forecasting and stock control for intermit-
tent demands. Operational Research Quarterly 23, 289
303.
Efron, B., Tibshirani, R.J., 1993. An Introduction to the
Bootstrap. Chapman and Hall, New York.
Feeney, G.J., Sherbrooke, C.C., 1966. The (s 1, s) inventory
policy under compound Poisson demand. Management Sci-
ence 12, 391411.
Gelders, L.F., van Looy, P.M., 1978. An inventory policy for
slow and fast movers in a petrochemical plant: A case study.
Journal of the Operational Research Society 29 (9), 867874.
Ghobbar, A.A., Friend, C.H., 2003. Evaluation of forecasting
methods for intermittent parts demand in the eld of aviation:
A predictive model. Computers and Operations Research 30,
20972114.
Heizer, J., Bender, B., 2001. Operations Management, 6th ed.
Prentice-Hall, New Jersey.
Kennedy, W.J., Patterson, J.W., Fredendall, L.D., 2002. An
overview of recent literature on spare parts inventories.
International Journal of Production Economics 76, 201
215.
Muckstadt, J., 2005. Analysis and algorithms for service parts
supply chains, Springer Series in Operations Research and
Financial Engineering.
Porras, E., 2005. Inventory Theory in Practice: Joint Replenish-
ments and Spare Parts Control, Tinbergen Institute Research
Series No. 363, Erasmus University Rotterdam, the
Netherlands.
Schultz, C.R., 1987. Forecasting and inventory control for
sporadic demand under periodic review. Journal of the
Operational Research Society 37, 303308.
Silver, E.A., Ho, Chung-Mei, Deemer, R.L., 1971. Cost mini-
mising inventory control of items having a special type of
erratic demand pattern. INFOR 9, pp. 198219.
Silver, E.A., Pyke, D.F., Peterson, R., 1998. Inventory Manage-
ment and Production Planning and Scheduling, third ed. John
Wiley & Sons, NY.
Strijbosch, L.W.G., Heuts, R.M.J., van der Schoot, E.H.M.,
2000. A combined forecast-inventory control procedure for
spare parts. Journal of the Operational Research Society 51,
11841192.
van der Heijden, M.C., Diks, E.B., de Kok, A.G., 1997.
Allocation policies in general multi-echelon distribution
systems with (R, S) order-up-to policies. International Journal
of Production Economics 38, 225244.
Willemain, T.R., Smart, C.N., Schwarz, H.F., 2004. A new
approach to forecasting intermittent demand for service parts
inventories. International Journal of Forecasting 20, 375
387.
Williams, T.M., 1984. Stock control with sporadic and slow-
moving demand. Journal of the Operational Research Society
35 (10), 939948.
ARTICLE IN PRESS

An Inventory Control System For Spare Parts at A Refiner

Uploaded by

Copyright:

Available Formats

An Inventory Control System For Spare Parts at A Refiner

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

An Inventory Control System For Spare Parts at A Refiner

Uploaded by

Copyright:

Available Formats

Production, Manufacturing and Logistics

An inventory control system for spare parts at a renery:

D and standard devi-

When we use the Poisson distribution to estimate

You might also like