European Journal of Operational Research 205 (2010) 313–324

Contents lists available at ScienceDirect

European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor

Production, Manufacturing and Logistics

Inventory management of multiple items with irregular demand: A case study

George Nenes, Sofia Panagiotidou, George Tagaras *
Department of Mechanical Engineering, Aristotle University of Thessaloniki, 54124 Thessaloniki, Greece

a r t i c l e i n f o a b s t r a c t

Article history: We present the case of a Greek commercial enterprise facing the problem of managing the inventories of
Received 28 November 2008 thousands of different items, supplied by more than 20 European and Asian manufacturers and sold to a
Accepted 27 December 2009 large number of different-type customers. A key feature of the problem is that the demand for the vast
Available online 11 January 2010
majority of items is intermittent and lumpy, thus not allowing the use of the usual normal or Poisson dis-
tributions. The paper describes the solutions given to several practical problems in the course of devel-
Keywords: oping an easy-to-use yet effective and all-encompassing inventory control system. Emphasis is placed on
Supply chain management
the accurate modeling of demand by means of a gamma distribution with a probability mass at zero or a
Inventory
Case study
package Poisson distribution for very-slow-moving items. Using those models and simple quantitative
Gamma distribution tools we develop an efficient procedure for approximate but quite accurate determination of the base
Periodic review stock levels that achieve the desired fill rates in the proposed periodic review system. We briefly describe
Base stock the computerized implementation of the new system and the very encouraging results.
Ó 2010 Elsevier B.V. All rights reserved.

1. Introduction standard textbook models as approximations in cases where the

Inventory management has been recognized as one of the most
important functions of industrial and commercial enterprises,
which often has a great impact on their overall performance. The
typical trade-off is between high holding and obsolescence costs
of excessive stock on one hand and poor service and high shortage
costs resulting from low inventory levels on the other. The desired
solution is a suitable inventory control policy that will guarantee a
satisfactory service level without keeping unnecessarily large
inventories that are costly and difficult to handle.
Inventory theory is probably the most thoroughly researched
area of production and operations management. However,
although almost all large companies and many small and med-
ium-sized enterprises increasingly try to apply scientific methods
for better managing their inventories, the use of these methods
is often limited to some basic tools like the computation of
economic order quantities and rough approximations of reorder
points or base stocks for achieving target service levels. The wide-
spread application of more elaborate and appropriate methods for
inventory control in practice is hindered by a number of factors
including most notably the following:
 The standard textbook material on inventory control is based on
the assumption that lead time demand follows a normal or Pois-
son distribution. This is often not the case in practice. Using the
* Corresponding author. Tel.: +30 2310996062.
E-mail addresses: gnenes@auth.gr (G. Nenes), span@auth.gr (S. Panagiotidou),
tagaras@auth.gr (G. Tagaras).
departures from the assumed normal or Poisson distribution are
serious leads to very unsatisfactory results.
 The standard textbook treatment of inventory management
either completely ignores or provides very vague guidelines
about the issues and complications that typically arise in prac-
tice due to the large number of suppliers and items that must
be coordinated, incomplete or suspicious data, lead time vari-
ability, possibility of late order placement or emergency replen-
ishment, etc.
The above reservations reflect precisely the skepticism at the
top management of RODA s.a., the largest distributor of castors
and wheels in Greece, when they asked us to assist them with
the rationalization and computerization of their inventory control
system. In a few words, the challenge that we faced and served as
the main motivation for our work was to bridge the gap between
theory and the practical peculiarities of a specific business environ-
ment so as to deliver a user-friendly and flexible computerized
decision support system for the management of thousands of stock
keeping units (SKUs), which the company orders and receives from
more than 20 suppliers located in Europe and China. A key and dis-
tinguishing feature of the problem is that the demand for the vast
majority of the SKUs is highly variable and irregular, thus defying
the usual assumptions of normal or Poisson distribution.
In addition to addressing a specific company’s problem, our
work touched upon issues and technical problems of broader inter-
est that, in our opinion, are worth communicating to the scientific
community. The most interesting of these issues, problems, and

0377-2217/$ - see front matter Ó 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2009.12.022
314 G. Nenes et al. / European Journal of Operational Research 205 (2010) 313–324
findings are summarized in the present paper, which we view as
making a twofold contribution:
(i) At the managerial level, it provides a ‘‘complete” account
(subject to journal space limitations) of the solutions given
to several practical problems in the course of developing
an easy-to-use yet effective and all-encompassing inventory
control system. Some of these solutions may be directly or
indirectly usable in similar situations.
(ii) At a more technical level, the paper describes the details of
modeling intermittent and lumpy demand for a wide range
of average volumes (both for slow-moving and fast-moving
items), as well as an efficient procedure for approximate
but quite accurate determination of appropriate base stocks
and estimation of fill rates and other important operating
characteristics in the proposed periodic review system.
The next section contains a brief literature review in two inter-
secting directions: (a) case studies of inventory control systems
and (b) modeling of intermittent and lumpy demand for stock con-
trol purposes. Then, Section 3 describes the problem setting: the
company, its products and suppliers, and the empirical inventory
control and ordering policy that the company used to apply for
the replenishment of its stock. Section 4 is devoted to the analysis
and modeling of demand for the various types of SKUs. The section
that follows describes the general features of the periodic review
base stock system that was proposed with the aim of introducing
more structure and streamlining the procurement operations. Sec-
tion 6 presents, explains and validates the formulas for the compu-
tation of critical operating characteristics of the proposed system.
Section 7 presents the determination of the appropriate base
stocks for each class of SKUs and Section 8 describes briefly the
decision support system that was developed for classifying the
SKUs into different classes and then computing the respective rec-
ommended base stocks. The results of the new system develop-
ment and implementation are discussed in the concluding section.
2. Literature review
The scientific literature in the area of inventory theory is so huge
that any attempt to provide a comprehensive coverage of even a
small part of it in a few journal pages would be doomed to failure.
However, reports of inventory control practice in the form of case
studies published in academic journals are much more limited. In
the following paragraphs we first refer to a few selected papers,
which present case studies having common general features with
the problem that we had to address. Then, we present publications
that deal specifically with the issue of modeling intermittent and
lumpy demand, placing emphasis on articles motivated by practi-
cal applications and focusing on specific variations of the Poisson
and gamma distributions, which have been employed in the past
and are also used in the analysis of our case in later sections.
2.1. Case studies of inventory control systems
During the last decades, a number of inventory control case
studies have appear in the academic literature, reporting applica-
tions mostly in the electronics, chemical, and automotive indus-
tries. One of the earliest such publications is by Gelders and van
Looy (1978), who present various inventory policies both for
slow-moving and fast-moving items in a petrochemical plant with
around 22,500 SKUs.
Vereecke and Verstraeten (1994) describe an algorithm for the
implementation of a computerized inventory management system
for spare parts in a large chemical plant, located in Belgium. The
spare parts inventory contains about 34,000 different types of
items, 90% of which have demand frequency lower than 4 times
a year.
More recently, Aronis et al. (2004) develop a Bayesian method-
ology to obtain more accurate forecasts for the demand of spare
parts of electronic equipment and then apply it to obtain the
appropriate values of the parameter S of an (S  1, S) inventory sys-
tem for these parts. Kukreja and Schmidt (2005) present the case of
a large utility company having 29 power generating plants in five
American south-eastern states. The inventory under study consists
of low-usage but expensive items with a lumpy demand pattern.
They develop and propose a continuous review inventory system
with transhipments in case of stockouts. Syntetos and Boylan
(2006) use simulation to measure the effectiveness of a periodic re-
view order-up-to inventory system for controlling the stock of
items with intermittent demand. They use real data from the auto-
motive industry and evaluate the system with both statistical and
economic criteria. Porras and Dekker (2008) compare different
reorder point methods for effective spare parts inventory control
at a large oil refinery in the Netherlands. They note that it is diffi-
cult to devise good strategies for the management of spare parts
because they are typically slow-moving with highly stochastic
and erratic demands.
2.2. Modeling of irregular demand
A common finding in most inventory control case studies is that
the demand of specific SKUs cannot always be modelled accurately
by means of the normal or Poisson distribution, as is usually as-
sumed in standard textbooks, because it is often erratic, intermit-
tent and/or lumpy. In addition to the references in the preceding
sub-section, see also the papers of Williams (1984) and Willemain
et al. (2004). Dolgui and Pashkevich (2008) report that 37.3% of the
approximately 145 million SKUs that the UK Royal Air Force (RAF)
kept in stock at the beginning of year 2000 had fewer than 10 de-
mand transactions over an observation period of 6 years.
Before proceeding further it is useful to clarify the terms erratic,
intermittent, sporadic, lumpy, slow-moving, irregular, because they
are not always used unequivocally. According to the classification
proposed by Ghobbar and Friend (2002), the demand is character-
ized as intermittent when it appears randomly with many time
periods having no demand; an erratic demand pattern is character-
ized by highly variable demand size; lumpy demand is both inter-
mittent and erratic; slow-moving items have intermittent demand
with each demand size equal to one item or very few items. How-
ever, the terms sporadic and intermittent are frequently used inter-
changeably (Dunsmuir and Snyder, 1989) and the same can be said
for the terms lumpy and irregular (Regattieri et al., 2005), as well as
for the terms intermittent and irregular (Willemain et al., 2004). In
general, irregular demand has a broader interpretation, encompass-
ing essentially all demand types that cannot be expressed by
means of the usual normal and Poisson distributions; the title of
the present paper reflects this broader definition.
The difficulties arising from the sporadic and erratic nature of
the demand, especially of spare parts, are often addressed in the
inventory literature with a focus on suitable forecasting methods.
Examples of such publications include Croston (1972), Rao
(1973), Willemain et al. (1994), Johnston and Boylan (1996),
Bartezzaghi et al. (1999), Ghobbar and Friend (2003), Regattieri
et al. (2005), and Syntetos and Boylan (2005). These works do
not explicitly connect the forecasting methods to any specific
inventory model. There are only a few papers that link forecasting
methods directly to inventory control models. Specifically, Watson
(1987) studies the lumpy demand case and examines the interac-
tions between demand forecasting and continuous review reorder-
ing subsystems, while Syntetos and Boylan (2006) and Teunter and
 G. Nenes et al. / European Journal of Operational Research 205 (2010) 313–324
Sani (2008) use forecasting methods in conjunction with a periodic
review order-up-to inventory control policy.
Another set of relevant publications discuss the suitable model-
ing of irregular demand with a view to applying the findings to the
development of effective inventory models. When demand is inter-
mittent but unitary the typical assumption is that it follows a Pois-
son distribution or equivalently that the time until a demand
appears is an exponentially distributed random variable. In such
cases, the usual inventory control policies are of the (S  1, S) type.
When, however, demand is generally or may be non-unitary, as is of-
ten the case in real applications, a more appropriate demand distri-
bution is the compound Poisson which can be viewed as a series of
customers with Poisson arrivals, where the quantity demanded by a
customer is an independently distributed random variable. Feeney
and Sherbrooke (1966) propose an (S  1, S) ordering policy under
compound Poisson demand. The compound Poisson distribution
has also been used by Cheung (1996), who allows the batch resupply
times to depend on the batch size, Thompstone and Silver (1975),
who study the coordination of multiple items replenishments, and
Forsberg (1995) and Zhao (2008) in a supply chain context.
A special case of the compound Poisson is the package Poisson
distribution, proposed by Friend (1960), where the demand size
of all customers is constant (a package). This distribution is used
by Williams (1984), who develops stock control systems appropri-
ate for items with various types of irregular demand. Vereecke and
Verstraeten (1994) conduct empirical research by means of com-
puter simulation using demand data for many slow-moving SKUs
in a large chemical plant. They arrive at the conclusion that the
package Poisson probability distribution fits well the observed de-
mand patterns.
When the magnitude of the demand is large but neither sym-
metric nor too erratic, the usual alternative to the normal distribu-
tion is the gamma distribution (Silver and Robb, 2008). This
distribution has been repeatedly reported as providing a good rep-
resentation of the demand in real cases; see, for example, Burgin
and Wild (1967) and Strijbosch et al. (2002). The gamma probabil-
ity distribution has been used in the context of intermittent/irreg-
ular demand by Dunsmuir and Snyder (1989) and Segerstedt
(1994). Specifically, Dunsmuir and Snyder (1989) propose a contin-
uous review inventory system for determining reorder levels keep-
ing a specified customer service level. The distinguishing feature of
their model is the use of a gamma distribution with a spike at zero
to represent the relative frequency of periods with no demand.
This mixture of a gamma density with a probability mass at zero
is a promising candidate for satisfactory modelling of intermittent
but not slow-moving demand. Janssen et al. (1998) extend the
work of Dunsmuir and Snyder (1989) to a periodic review inven-
tory system with general demand distribution in a time period
and taking the inventory undershoot into account. Segerstedt
(1994) presents a periodic review model for inventory control with
variable lead times and intermittent demand assuming that the
times between demands and the size of each demand are gamma
distributed random variables. The approach is different from ‘‘typ-
ical” inventory models as no reorder point is explicitly computed; a
replenishment order is placed when the probability of a shortage
exceeds the required probability for no shortage.
Many other distribution types have been proposed in the past for
modelling irregular lead time demand but a complete account is be-
yond the scope of this brief and focused literature review. The inter-
ested reader may refer to Silver et al. (1998) for additional references.
3. Problem description
RODA s.a. is a small company based in Thessaloniki, Greece. It
was established in 1948 as a family business but has evolved into
315
the largest distributor of castors and wheels in Greece. It currently
employs 34 people in two business units: (a) institutional castors
manufacturing and (b) commercial. This work focuses on the com-
mercial business unit, which procures, stores, assembles, sells, and
distributes close to 10,000 SKUs, ranging in complexity from small
and inexpensive bolts to large heavy-duty wheels for industrial
applications. Most of these SKUs are variations of the main product
types, which are assembled in house from basic components and/
or intermediate products that are provided by the company’s
suppliers.
RODA’s commercial business unit procures approximately 3000
components and intermediate and final products from 28 suppli-
ers, which are manufacturing companies located in several Euro-
pean countries and in China. Twenty-three suppliers are used on
a regular basis, while five suppliers are used very infrequently
and only for customized products. One of the suppliers is RODA’s
own production plant, which is the sole manufacturing facility of
institutional castors in Greece. The distribution of total annual va-
lue across the SKUs of every single supplier follows the typical ABC
analysis pattern, with many SKUs showing absolutely no demand/
consumption in one and even two or more years.
The ordering policy of the procurement department is supplier-
dependent. At the time that the project was about to start, that is in
early 2007, the procurement department placed orders on a con-
tinuous basis following empirical rules about implicit safety stock
targets, knowing however that these orders would be delivered
periodically, provided a minimum order size would be accumu-
lated. To give a specific example, a certain German supplier would
assemble, load and send off an order to Greece by truck on Thurs-
days or Fridays if the total weight exceeded 300 kilograms. In prac-
tice, that happened every other week with a few exceptions of
more or less frequent deliveries. The truck would arrive at RODA’s
warehouse after 5 or 6 days containing only the items that were
available for delivery at the time that the order was assembled at
the supplier, i.e. 1 or 2 days before the truck was actually loaded.
Thus, the effective lead time depended on the availability of the re-
quested item at the supplier, which in turn depends on its utiliza-
tion, as well as on the time distance between the placement of the
specific order and the truck departure. In short, the effective lead
time ranged from 8 days for some fast-moving items when their
orders were placed just before the order assembly at the supplier
to a few weeks for slow-moving items that the supplier would
manufacture to order.
Furthermore, the suppliers differ substantially in terms of their
delivery frequency and transportation time. For example, one of
the major Chinese suppliers ships an order once a month and it
takes about two months for the shipment to arrive in Greece, clear
customs and become available for sale, while suppliers of low-vol-
ume make-to-order specialized items may make as few as only two
shipments per year.
The demand side is equally complex, as the customer base in-
cludes many small customers making frequent or infrequent pur-
chases of small quantities and expect immediate service along
with a few large ‘‘institutional” customers (hospitals, super mar-
kets, and large manufacturers) that typically place large orders
but usually allow a reasonably long delivery time. These ‘‘planned”
orders, which sometimes take the form of a large quantity that
RODA must deliver to the customer during an entire year in six
equal bimonthly installments, do not pose a problem to the inven-
tory system as long as the time window between order placement
and requested delivery time exceeds the maximum effective lead
time for the specific item, which is usually the case; this type of de-
mand can be easily satisfied by placing special orders to the respec-
tive supplier (in addition to the regular orders) with sufficient
advance notice, thus making sure that the desired quantity will
be available exactly when the customer needs it.
316 G. Nenes et al. / European Journal of Operational Research 205 (2010) 313–324

4. Demand modeling 60 weeks exhibiting zero demand, while the histogram of the 44
non-zero weekly demands reveals positive skewness. Taking into
Exceptionally large sporadic demands from large customers dis- account on one hand this type of demand pattern, which repeats
tort the picture of the demand distribution and render demand itself in the vast majority of A items, and on the other hand the rel-
modeling a nightmare. As was explained at the end of the previous evant academic literature, most notably the work of Dunsmuir and
section, such singularities can and must be excluded from the reg- Snyder (1989), it is natural to consider modeling the weekly de-
ular demand time series if they correspond to ‘‘planned” demand. mand X using a gamma distribution with a probability mass at
That was the case at RODA, where the raw data set covering the zero. The probability density function (pdf) of the gamma distribu-
two-year period 2005–2006 had to be cleaned up very carefully tion for the strictly positive weekly demand X þ ¼ ðXjX > 0Þ is
by removing all instances of ‘‘planned” demand. This required a
lot of work on the part of the company, going deeply into the re- kc xc1 kx
fX þ ðxÞ ¼ e ; x > 0;
cords of every single unusually large demand transaction to find CðcÞ
out whether the demand was indeed ‘‘planned” or simply an unex- R1
where CðcÞ ¼ 0 eu uc1 du, and k and c are the scale and shape
pected large demand without prior notice. The good news was that
parameters, respectively. The mean l and the variance r2 of the
this exercise not only removed some extreme data points from sev-
above density function of X þ are l ¼ c=k and r2 ¼ c=k2 , respec-
eral demand streams, it also led to the complete exclusion of a sig-
tively. The probability of non-zero demand in a week, p, constitutes
nificant number of the A items that were examined first (in terms
the third demand model parameter, which along with k and c fully
of annual value) because their entire demand was planned and
determine the distribution of X.
consequently they could be managed without keeping any inven-
The gamma distribution is very convenient because of its flexi-
tory at all. In fact, that is exactly what is still happening with these
bility, achieved by choosing appropriate k and c values, and its
items; RODA’s warehouse serves as a cross-docking facility for
highly useful reproductive property: if the demand in a time unit
them. The bad news was that even after that careful ‘‘cleanup”
is a random variable following a gamma distribution with param-
the demand time series and histograms for the vast majority of
eters k and c, then the total demand in n time units also follows a
the many remaining A items did not appear to provide any support
gamma distribution with parameters k and cn ¼ nc. Most impor-
for the usual assumptions of normal or Poisson demand distribu-
tantly, the gamma distribution in our case describes with satisfac-
tions. Furthermore, the picture was even bleaker for the B and C
tory accuracy the demand of most A items, according to the chi-
items.
square goodness-of-fit tests, which we performed using the avail-
In the remainder of this section we explain how we addressed
able historical data.
the problem of modeling the irregular unplanned demand of the
The specific procedure that was followed for the estimation of
various SKUs, making a distinction between ‘‘faster-moving” items,
the parameters p, k, c, of the demand model of every item is illus-
where the demand is expressed by means of the gamma distribu-
trated below through the example of W200. First, the probability of
tion (Section 4.1) and ‘‘slower-moving” items, where a variation of
positive demand in a week is estimated from the available two-
the Poisson distribution yields satisfactory results (Section 4.2).
year data as the proportion of weeks where demand was not zero:
This distinction is applicable to all SKUs, irrespective and indepen-
p = 44/104 = 0.423 (for the sake of simplicity there is no distinction
dent of their classification as A, B or C; although A items tend to be-
in the notation of true unknown parameters and their point esti-
long to the faster-moving category, many of the slower-moving
mates, which are used in all the computations that follow). Then,
SKUs are also A items because of their high unit purchase cost.
the actual X þ values are used to estimate l and r2 of the assumed
The demand data analysis was eventually performed for all
gamma distribution and then the values of k and c are computed
SKUs from all suppliers because it was decided to develop a univer-
from k ¼ l=r2 and c ¼ ðl=rÞ2 .
sal system capable of managing the total stock of the company
A chi-square goodness-of-fit test is then performed to ensure
with no exceptions. It is worth noting that the time unit was either
that the gamma distribution is an acceptable representation of
a week or a month, depending on the suitable time scale of the
the behavior of the random variable X þ . For W200, the 44 positive
respective supplier. For example, the demand data for the SKUs
weekly demand values yield l ¼ 40:14 and r ¼ 55:42, which in
supplied from China by sea was analyzed on a monthly basis (24
turn lead to k ¼ 0:013 and c = 0.52. The minimum significance level
data points in 2005–2006), while the demand data for the rela-
for rejection of the hypothesis that X þ follows a gamma distribu-
tively fast-moving SKUs supplied from Europe by truck every one
tion is 0.66.
or two weeks was analyzed on a weekly basis (104 data points in
As it will become evident in Section 6, what is eventually neces-
2005–2006). For ease of exposition, though, we mostly refer to
sary for the inventory system analysis is the distribution of the to-
‘‘weekly demand” rather than ‘‘demand per time unit.”
tal demand in multiple weeks (lead time + review period). This
It must be noted that the demand data analysis described below
demand is the convolution of weekly demands. Its determination
implicitly assumes that demand is stationary. Although in reality
is facilitated by the reproductive property of the gamma distribu-
the demand for most items exhibits a positive trend (without sea-
tion but at the same it is complicated by the existence of the prob-
sonality), the rate of increase is low enough to be neglected for
ability 1  p of no demand in a given week.
simplicity, without severely affecting the accuracy and effective-
ness of the analysis. An update of the estimated distribution
parameters every 6 months is deemed sufficient for taking into ac- 4.2. Slower-moving items
count any sizeable changes in the demand pattern.
There are many SKUs, especially in the B and C classes of the
ABC classification, exhibiting so sporadic and low demand that
4.1. Faster-moving items the gamma distribution cannot provide a good fit with any choice
of parameters. The Poisson distribution is a suitable choice in only
Fig. 1 shows the time series and the histogram of the weekly de- a few of these cases, because of the following peculiarity: since
mand in 2005–2006 for a typical relatively fast-moving item sup- most SKUs are castors and wheels, the customers tend to order
plied from Germany, which we denote by the fictitious code them in pairs. Therefore, the demand distribution looks often like
number W200 and will serve as an illustrative example here and the observed frequencies for item TE320B in the 24 months of
in later sections. It is clear that the demand is intermittent with 2005–2006, which is reproduced in Table 1.
 G. Nenes et al. / European Journal of Operational Research 205 (2010) 313–324 317

300

250

demand 200

150

100

50

0
2005 2006
time

70

60

50
frequency

40

30

20

10

0
0 1 - 32 33 - 64 65 - 96 97 - 128 129 - 160 161 - 192 193 - 224 225 - 256
weekly demand

Fig. 1. Time series and histogram of weekly demand of W200 in 2005–2006.

Table 1 5. The proposed inventory management system

Observed monthly demands for item TE320B in 2005–2006.

Monthly demand 0 2 4 6 8 In order to streamline and standardize the procurement proce-

Frequency 5 7 5 4
The cases where the demand distribution has the appearance of
Table 1 are modeled through the package Poisson distribution,
namely a Poisson distribution for demand occurrence combined
with a constant demand size f, which is called package. In the case
of RODA the number of units in the package is f = 2. Therefore, the
weekly or monthly demand quantities are converted to numbers of
packages (pairs) and the resulting frequency distribution is tested
for goodness-of-fit to the Poisson distribution. The outcomes of
these tests showed that the package Poisson distribution expresses
the demand pattern with satisfactory accuracy for many items.
Note that the regular Poisson distribution can be viewed as a spe-
cial case of the package Poisson with f = 1.
Finally, there is an even larger number of SKUs with so low and
infrequent demand that the available data are not sufficient for
reliable statistical analysis and characterization of their demand.
The practical threshold for declaring items as having insufficient
data was set to a total demand of less than 15 units and/or less
than 7 periods with non-zero demand in 2 years. The inventory
policy for those SKUs is based on simple rules as explained in
Section 7 rather than on an explicit analysis of their very sparse
demand pattern.
3 dures in a complex real system, it is clearly advisable to use the
same type of inventory management policy for all the SKUs and
all the suppliers, if that is possible. After extensive consultation
with the top management of the company and taking seriously
into account current practices and technical constraints, it was
decided to adopt and universally implement the periodic review
system (R, S), where R stands for the review period while S is the
base stock. Thus, the ordering policy for all SKUs is of the common
order-up-to-S type; at every review instance the order quantity for
any item is S – IP, where IP is the inventory position of that item,
namely the stock that is either physically available (negative in
the case of a temporary shortage/backorder) or has been previously
ordered but not yet received. In effect, the order quantity is equal
to the demand that has been recorded since the previous review
point.
The operation of the (R, S) system can be viewed as a succession
of cycles with generally different durations and timing for different
suppliers, since the review period R is not the same for all suppli-
ers. For any specific supplier with given R the duration of each cy-
cle is R. Its beginning and ending points are the times where orders
are received from that supplier. It is possible, though, that no order
is placed with a supplier at some review instance, because no de-
mand for its products was expressed in the immediately preceding
time interval R. In that case, which is not uncommon for some sup-
pliers of specialized low-demand castors and wheels, the review
318 G. Nenes et al. / European Journal of Operational Research 205 (2010) 313–324

cycle is still defined as having the same length R and ends at the 6.1. Expected stock on hand
time when an order would have been received had it been placed.
The universal application of the (R, S) system in a company with In typical (R, S) systems an inventory cycle lasts R time units,
many suppliers and thousands of SKUs presents several practical starting just after an order is received (time t1 in Fig. 2) and ending
challenges related on one hand to the frequency and timing of or- just before the next order is received (time t 2 in Fig. 2). The net
ders and deliveries and on the other hand to the coordination of stock at the beginning ðNSB Þ and at the end ðNSE Þ of each inventory
multiple items per supplier, differing in availability and effective cycle may be either positive or negative, generating the three pos-
lead time as explained in Section 3. Thus, although in theory the re- sible cases illustrated in Fig. 2.
view period R is a decision variable, which can be optimized to- Since the actual expected stock on hand, E(OH), depends on
gether with the other decision variable S through an appropriate whether NSB and NSE are positive or not, it is calculated differently
cost-minimization model, for the sake of smoothness of transition in each of the three cases of Fig. 2. Specifically, E(OH) is the average
to the new system it was decided to maintain the same order/ship- of NSB and NSE when they are both positive and thus coincide with
ment frequencies as before. The implication of this choice is that the available inventory (case A), but it is generally different from
for each supplier the value of R is taken as given, i.e., it is deter- that when there is a stockout during part of the cycle (case B)
mined by the company’s management based on practical and eco- and it reduces to zero when the cycle begins and ends without
nomic considerations such as vehicle or ship itineraries, minimum any inventory available on hand (case C). Consequently, E(OH)
economic shipment quantities and the structure of transportation must be derived taking into account and weighing appropriately
costs. Then, given the timing of shipment, the composition of the the above cases. Assuming that demand is continuous and strictly
shipment (truckload or container) is determined taking into ac- positive in every period (p = 1), the final expression for E(OH) is
count the outstanding orders, which are placed according to the Z S Z SxL
1
new base stock system. EðOHÞ ¼ ðS  xL ÞfR ðxR ÞfL ðxL ÞdxR dxL
Even though the frequencies of deliveries remained unchanged, 2 0 0
Z S
the ordering system was streamlined considerably by specifying a
þ ðS  xRþL ÞfRþL ðxRþL ÞdxRþL
cut-off time point for the placement of orders at each supplier, suf- 0
ficiently ahead of the scheduled departure of the truck from the Z S Z 1 2
supplier’s premises, to ensure that the assumed lead times will þ ðS  xL ÞfR ðxR ÞfL ðxL ÞdxR dxL
0 SxL
not be exceeded. Consider, for example, the German supplier men- Z Z 
S 1
tioned in Section 3 that sends off a shipment every other Thursday
xR fR ðxR ÞfL ðxL ÞdxR dxL ; ð1Þ
or Friday. By placing the orders not later than a week in advance of 0 SxL
the departure and taking into account the 5–6 days transportation
where xR ; xL and xRþL denote the demand in R, L and R + L time units,
time, it is practically certain that all the ordered items will be re-
with density functions fR ðxR Þ; f L ðxL Þ and fRþL ðxRþL Þ, respectively. The
ceived after a lead time L = 2 weeks, except for those specialized
derivation of (1) is presented in detail in Appendix A.
items, which are manufactured to order (MTO) and typically arrive
It is evident from (1) that the exact computation of E(OH) is
with the next shipment two weeks later (L = 4 weeks).
cumbersome, especially considering that in our case the probabil-
The new system imposes more discipline in the ordering pro-
cess and enforces explicit negotiation and agreement with the sup-
pliers about the availability of the various SKUs, which in turn
specifies the realistic lead times that will be used in the analysis Case A
NSB
and determination of the base stock levels S for all items. In addi-
tion, it provides a window of flexibility and a cushion against
shortages at the supplier by still allowing orders to be placed be- NSE
fore the designated order time so as to give the supplier a better
chance to respond without delay, especially to orders of MTO t
items.
NSB Case B

6. Operating characteristics

In order to analyze and evaluate the effectiveness of the pro-

t
posed inventory management system and determine the appropri-
NSE
ate value of the base stock S for every item, it is necessary to be
able to compute the following key operating characteristics as
functions of S:
Case C
 E(OH): expected stock on hand;
 E(SC): expected shortage per cycle; NSB t
 b: fill rate (percentage of demand satisfied immediately from
available stock).

The computation of E(OH), E(SC) and b is described below in NSE

some detail for the case of faster-moving items with gamma distri-
bution of positive demand. When demand is modelled with the aid
of Poisson or package Poisson distribution the computation of L
these operating characteristics is standard, with only a straightfor- t1 R t2
ward adjustment to take into account the package size in the latter
case; hence the details are omitted. Fig. 2. Alternative cases of inventory cycles.
 G. Nenes et al. / European Journal of Operational Research 205 (2010) 313–324 319

Table 2
Expected stock on hand and expected shortage for item W200 for selected values of S.

b S EðOHA Þ EðOHN Þ E(OH) ES ðOHÞ EðSC B Þ EðSC E Þ E(SC)

0.80 161 120.1 110.1 115.1 115.77 ± .32 3.2 10.0 6.8
0.90 223 176.8 172.1 174.4 174.49 ± .36 1.4 4.8 3.4
0.95 283 234.4 232.1 233.2 233.05 ± .40 .6 2.3 1.7
0.99 420 369.5 369.1 369.3 369.03 ± .43 .1 .4 .3

ity 1  p of no demand in a time unit is not negligible for most

R + L  n time units). Consequently, the expressions for EðOHA Þ
SKUs (p < 1). We therefore prefer to resort to an efficient approxi-
and EðOHN Þ are modified and generalized for any arbitrary R and
mation, which is based on the usual definition of the expected
L as follows:
stock on hand in an (R, S) system as the average of the expected
beginning and ending inventory of a cycle. Since the beginning RþL  
X RþL n
inventory is the ending inventory of the previous cycle plus the EðOHA Þ ¼ ð1  pÞRþL S þ p ð1  pÞRþLn
quantity that is received, the expected beginning inventory equals n¼1
n
the expected ending inventory plus the expected order quantity. Z S
DR
The latter coincides with the average demand in a cycle, which is  ðS  xn Þfn ðxn Þdxn þ
; ð3Þ
0 2
DR, where D is the average weekly demand. Consequently, RþL  
X RþL n
E(OH) = E(Ending Inventory) + DR/2. Hence, evaluation of E(OH) is EðOHN Þ ¼ ð1  pÞRþL S þ p ð1  pÞRþLn
based on the computation of the expected inventory at the end n¼1
n
Z 1
of a cycle. DR
The expected ending inventory in a cycle can be expressed
 ðS  xn Þfn ðxn Þdxn þ
0 2
either as the expected ending on-hand inventory EðOHE Þ or as the RþL 
X 
RþL n DR
expected ending net stock EðNSE Þ. Clearly, using EðOHE Þ for the ¼ p ð1  pÞRþLn ½S  Eðxn Þ þ : ð4Þ
n 2
computation of the expected ending inventory results in overesti- n¼0

mation of E(OH), while using EðNSE Þ leads to underestimation of If the positive demand in a time unit is expressed as a random
E(OH). One simple solution to this problem is to evaluate E(OH) variable following a gamma distribution with parameters k and c,
both ways and then use the average of the two values. Extensive then the reproductive property of the gamma distribution implies
simulations have shown that this heuristic approximation yields that fn ðxn Þ is a gamma pdf with parameters k and nc.
results very close to the actual (simulated) values of E(OH) and For an illustration of the above formulas and their accuracy,
consequently it may be adopted for practical purposes. consider again item W200 having average weekly demand
More specifically, assuming for the moment that demand is D ¼ pl ¼ 0:423  ð40:14Þ ¼ 16:98, lead time L = 2 weeks and review
continuous and strictly positive in every period (p = 1), the ex- period R = 2 weeks. Table 2 shows selected values of S (resulting in
pected ending available (on-hand) inventory in a cycle is fill rate values 0.99, 0.95, 0.90, 0.80) and the corresponding values
Z S of EðOHA Þ; EðOHN Þ and E(OH), computed from (2)–(4). The same ta-
EðOHE Þ ¼ ðS  xRþL ÞfRþL ðxRþL ÞdxRþL : ble also contains ES ðOHÞ, which are the simulation estimates of
0 E(OH) in the form of 95% confidence intervals. The direct compar-
Similarly, the expected ending net stock is ison of E(OH) against ES ðOHÞ offers an example of the satisfactory
Z accuracy of the proposed approximations.
1
EðNSE Þ ¼ ðS  xRþL ÞfRþL ðxRþL ÞdxRþL : If the demand per time unit follows a package Poisson distribu-
0 tion with package size f = 2 (or simply the Poisson distribution with
f = 1), then the expected stock on hand is computed again from (2)
The expected stock on hand using the ending available inventory
but with
is
Z S X
S
EðOHA Þ ¼ ðS  xRþL ÞfRþL ðxRþL ÞdxRþL þ DR=2: EðOHA Þ ¼ f ðS  xRþL ÞPðxRþL Þ þ fDR=2;
0 xRþL ¼0

The expected stock on hand using the ending net stock is EðOHN Þ ¼ f ½S  DðR þ LÞ þ fDR=2 ¼ f ½S  DR=2  DL;
Z 1
where Pðxn Þ is the Poisson distribution corresponding to n time
EðOHN Þ ¼ ðS  xRþL ÞfRþL ðxRþL ÞdxRþL þ DR=2: units.
0

Finally, E(OH) is computed as (the positive part of) the average

of EðOHA Þ; EðOHN Þ: 6.2. Expected shortage per cycle and fill rate
 þ
EðOHA Þ þ EðOHN Þ A usual assumption in the standard textbook analysis of peri-
EðOHÞ ¼ : ð2Þ odic review systems is that after receipt of an order all outstanding
2
backorders are satisfied and consequently the probability of a
Taking now into account that the probability 1  p of no shortage at the beginning of a cycle is negligible (Silver et al.,
demand in a time unit is generally not negligible (p < 1), the total 1998). This approximation works well when the total demand in
demand in R + L time units has to be expressed as a mixture of a R + L is regular enough to be modeled as a normally distributed
probability mass ð1  pÞRþL at zero and R + L density functions random variable, but it is not accurate enough when the demand
 
RþL n is highly irregular and intermittent as is the case at hand. There-
fn ðxn Þ with respective probabilities p ð1  pÞRþLn ; n ¼
n fore, the expected shortage per cycle, E(SC), is computed here as
1; 2; . . . ; R þ L, where fn ðxn Þ stands for the distribution of suggested by Zipkin (1986), by subtracting the expected shortage
positive demand xn in n time units (and no demand in the other in the beginning of the cycle, EðSC B Þ, from the expected shortage
320 G. Nenes et al. / European Journal of Operational Research 205 (2010) 313–324

at the end of the cycle EðSC E Þ. When p = 1 the latter is computed estimate; in the case of RODA it was practically impossible to ob-
from tain any reliable estimates, considering the variable importance
Z 1 of the customers, the vast number of SKUs, and the partial or full
EðSC E Þ ¼ ðxRþL  SÞfRþL ðxRþL ÞdxRþL ; substitutability between several SKUs, depending on the specific
S
customer requirements. Therefore, it was agreed to use a statistical
while EðSC B Þ is given by criterion for the choice of the base stocks, namely to find for each
Z 1 item the value of S that provides a target service level, which de-
EðSC B Þ ¼ ðxL  SÞfL ðxL ÞdxL : pends on the importance of the item.
S
The service level is expressed by the fill rate b, that is, the ex-
Thus, pected long-term fraction of demand that is satisfied immediately
Z 1 Z 1 from the available stock. The company first specifies the target fill
EðSCÞ ¼ ðxRþL  SÞfRþL ðxRþL ÞdxRþL  ðxL  SÞfL ðxL ÞdxL : rates for all items that are ordered to stock, by classifying them in
S S
four categories with respective target b values 0.99, 0.95, 0.90,
The general form of the above expressions for p 6 1 and arbi- 0.80, according to the perceived (mostly subjective) criticality of
trary R and L are as follows: a stockout. However, it has the flexibility to change the desired fill
! Z rate to any value within or even outside the (0.80, 0.99) region.
X
RþL RþL 1
EðSC E Þ ¼ pn ð1  pÞRþLn ðxn  SÞfn ðxn Þdxn ; Given a target fill rate and sufficient data to model the weekly
n¼1 n S or monthly demand as a random variable following either the gam-
! Z ma distribution with a probability mass at zero or the package
X
L L 1
EðSC B Þ ¼ pn ð1  pÞLn ðxn  SÞfn ðxn Þdxn ; Poisson distribution, it is easy to determine the base stock value
n¼1 n S
S that provides the required fill rate b through (5) and (6),
! Z
X
RþL RþL 1 respectively.
EðSCÞ ¼ pn ð1  pÞRþLn ðxn  SÞfn ðxn Þdxn When the available data are not sufficient to characterize the
n¼1 n S
! demand, according to the ad hoc criteria stated at the end of Sec-
X
L Z
L 1 tion 4, then simple rules are needed to provide guidelines for the
 pn ð1  pÞLn ðxn  SÞfn ðxn Þdxn choice of S. The heuristic approach that was followed to devise
n¼1 n S
such rules consists of two steps:
Since E(SC)/DR is the fraction of demand that is not satisfied di-
rectly from stock, the fill rate is computed from (i) Compute the values of S that result in b ¼ 0:80, 0.90, 0.95,
0.99 for a large number of SKUs with sufficient data for all
EðSCÞ
b¼1 : ð5Þ possible combinations of R and L; also record the respective
DR
mean demands in R þ L ðlRþL Þ.
Table 2 contains the values of EðSC E Þ; EðSC B Þ, E(SC) and b corre- (ii) Search for the functional relationships between lRþL and S,
sponding to selected base stocks S for item W200. Note that, espe- which best fit the results in step (i).
cially for lower values of S, the expected shortage at the beginning
of the cycle, after an order is received, is not negligible at all. Ignor- It turns out that a very good fit is achieved with relationships of
ing it would result in gross underestimation of the fill rate for given the simple form S ¼ AlBRþL , where the best values of the parameters
S. For example, EðSC B Þ ¼ 3:2 for S = 161; using EðSCÞ ¼ EðSC E Þ ¼ A and B depend on the values of b, R and L. Table 3 summarizes the
10:0 rather than E(SC) = 6.8 in (5) would lead to b ¼ 0:71 instead heuristic expressions for S when the appropriate time unit is a
of the correct b ¼ 0:80. week; a similar table has also been constructed for items managed
If the demand per time unit is package Poisson (f = 2) or Poisson on a monthly basis. The expressions of Table 3 are based on at least
(f = 1), then the expected shortage per cycle and the fill rate are 22 and up to 193 ðlRþL ; SÞ pairs for every b, R, L combination. Fig. 3
evaluated using the following expressions: shows the curves with the best fit for R = 2, L = 2 and b ¼ 0:80, 0.90,
X
1 0.95, and 0.99. It is important to note that the only required input
EðSC E Þ ¼ ðxRþL  SÞfPðxRþL Þ is the mean demand lRþL , which is always readily available. It is
xRþL ¼S also interesting to mention the resemblance of our heuristic ap-
X
1
proach to the power approximation method devised and proposed
EðSC B Þ ¼ ðxL  SÞfPðxL Þ
30 years ago by Ehrhardt (1979).
xL ¼S
The approximate relationships of Table 3 may also be used for
X
1 X
1
EðSCÞ ¼ ðxRþL  SÞfPðxRþL Þ  ðxL  SÞfPðxL Þ the determination of S values for C items that have sufficient de-
xRþL ¼S xL ¼S mand data but are not considered important enough to burden
EðSCÞ the system with the estimation of their demand parameters and
b¼1 : ð6Þ the computations needed to obtain S through the expressions of
fDR
Section 6. However, the company is cautioned that the actual fill
rate associated with the approximate base stock S may deviate
7. Determination of base stocks substantially from the prescribed one. Moreover, if S is computed
from such a relationship, there is no estimate of the expected stock
Having decided that the ordering/inventory policy will be of the on hand, which is an operating characteristic of great practical
(R, S) type for all suppliers and SKUs, with fixed supplier-depen- interest to the management of the company as it allows the esti-
dent review periods R as explained in Section 5, the remaining ele- mation of the average value of stock on hand.
ment for the full specification of the system is the determination of A final category of items, as far as the determination of S is con-
the base stock S for every single item. In theory, the appropriate cerned, consists of the following two sub-categories: (a) special
procedure for determining the optimal S in a periodic review SKUs for specific customers (including some A items), which RODA
(R, S) system with given review period R is to find the value of S manages explicitly on an order-by-order basis, and (b) SKUs that
that minimizes the total expected inventory holding and shortage may have been stocked in the past in anticipation of demand but
costs. In practice, though, the unit shortage cost is often too hard to the two-year historical data analysis shows that their demand
 G. Nenes et al. / European Journal of Operational Research 205 (2010) 313–324 321

Table 3
Simple rules for various combinations of fill rate, R, L (weeks).

Fill rate R = 1, L = 1 R = 1, L = 2 R = 2, L = 1 RþLP4

0.80 S ¼ 3:403lRþL S ¼ 3:403l 0:9645
RþL S ¼ 3:137l 0:962
RþL S ¼ 4:675l0:7785
RþL

0.90 S ¼ 4:275lRþL S ¼ 4:275l 0:9799

RþL S ¼ 4:078l 0:9739
RþL S ¼ 6:442l0:7659
RþL

0.95 S ¼ 5:509lRþL S ¼ 5:509l0:9762

RþL S ¼ 5:212l0:9735
RþL S ¼ 8:314l0:7525
RþL

0.99 S ¼ 7:670lRþL S ¼ 7:670l0:9857

RþL S ¼ 7:340l0:9834
RþL S ¼ 10:398l0:7785
RþL

2500
β = 0.99 S = 10.398 μ R+L 0.7785

2000 β = 0.95
β = 0.90
β = 0.80 S = 8.314 μ R+L 0.7525
1500
S = 6.442 μ R+L 0.7659
S

1000 S = 4.675 μ R+L 0.7785

500

0
0 200 400 600 800 1000 1200 1400
μ R+L

Fig. 3. SðlRþL Þ curves for items with R = 2 weeks and L = 2 weeks.

has been extremely infrequent, with no more than two demand in-
stances in 2 years. In both cases there is no point in keeping any
stock on hand and consequently the base stock for all those SKUs
is set equal to zero.
8. The decision support system
In order to assist the management of RODA in the determina-
tion of the appropriate base stocks for all SKUs we developed a
computer program for the efficient computation of S values taking
into account the required service levels for the customers (fill
rates) and the characteristics of the respective suppliers (order fre-
quencies and lead times). In addition, the program makes recom-
mendations about specific elements of the problem that became
evident during the discussions and the data analysis, such as pos-
sible irregularities in the historical demand data and excessive ser-
vice level requirements. The final result can be viewed as a simple
but flexible decision support system, the operation of which is de-
scribed briefly step-by-step in the following paragraphs.
Step 0. Data input
For every SKU, the program reads its code number, review per-
iod R, lead time L, target fill rate b, and the historical demand string
of up to 104 weeks or 24 months excluding all planned orders.
Step 1. Check for sufficiency of demand data
For every SKU the program checks whether the available
demand data is sufficient for further statistical analysis. If it is, it
continues to demand analysis. If it is not (total demand less than
15 units and/or less than 7 periods with non-zero demand), it
issues the warning ‘‘insufficient data” and then:
(ii) If there exist less than three non-zero demand periods,
then the program suggests S = 0.
Step 2. Demand analysis
A test of the available sufficient data is performed to determine
the goodness-of-fit to the gamma and Poisson distributions. Items
with average annual demand higher than 250 units are tested for
conformance to the gamma distribution only, items with average
annual demand not exceeding 15 units are tested for fitness to
the Poisson distribution (both the usual Poisson and the package
Poisson) and items in the intermediate demand range are tested
against both distribution types. In the rare cases where the data
do not fit any of these distributions, then the gamma distribution
is used for the faster-moving items (average annual demand of at
least 50 units) and the more suitable of the Poisson distributions
is used for the slower-moving items.
Step 3. Search for spurious demand data
It is important to check if the demand time series is spurious,
i.e., too different from those of most other items. The purpose of
this check is to identify any demand outliers, possibly correspond-
ing to large planned orders, which have not been removed from the
(unplanned) demand string as they should have. The following
specific rules for characterizing a demand string as suspect have
been derived empirically through extensive demand data analysis:
(i) The coefficient of variation of demand in R + L ðrRþL =lRþL Þ
is too high, namely higher than 2.5 when the time unit is a
week or higher than 1.5 when the time unit is a month.
(ii) The highest observed weekly (monthly) demand exceeds
30% (40%) of the total two-year demand.

(i) If there exist at least three non-zero demand periods, a
simple rule of the form S ¼ AlBRþL is used to compute
the base stock S for the given fill rate b. The program also
computes S for b ¼ 0:80, if the given (target) fill rate is
higher than 0.80 and recommends that base stock value
S to the user, with the rationale that a higher service level
is unnecessary for such slow-moving items.
If at least one of the above holds, then, in addition to computing
S for the specified target fill rate b, the program issues the message
‘‘not recommended to use FR > 0:8” and computes a recommended
S, which corresponds to b = 0.80.
Step 4. Computation of base stock and other characteristics
The program computes the lowest S that achieves the target fill
rate, as well as the values of the following demand and operating
322 G. Nenes et al. / European Journal of Operational Research 205 (2010) 313–324

Table 4
Sample program output; Poisson2 stands for package Poisson.

Code t.unit R L d.fitted d.used E(D) CV P(D > 0) S E(OH) FR

A19/25 Month 1 2 Gamma Gamma 31.1 1.4 .92 258 181.4 .95
A25/14 Month 1 2 Gamma Gamma 20.8 2.9 .42 425 374.2 .95
B18X7 Month 2 1 Poisson2 Poisson2 1.0 2.0 .29 8 6.1 .98
G125 Week 2 2 Insufficient data 2.7 8.6 .03 49 .95
G125 Not recommended to use FR > 0.8 29 .80
G160A Week 2 2 Insufficient data .1 9.1 .02 0 0.0 .00
G160N Week 2 2 Gamma Gamma 3.2 3.9 .25 109 99.5 .95
G160N Not recommended to use FR>0.8 57 48.4 .80
K390 Week 2 3 None Poisson2 .6 2.5 .18 8 5.8 .97

characteristics, which are of interest to the management of the
company:
 E(D): mean demand per time unit (month or week).
 CV: coefficient of variation of demand per time unit.
 P(D > 0): proportion of time units with non-zero demand.
 E(OH): average stock on-hand; not computed when S is
derived by the simple rules.
mended S values for 1420 SKUs, 814 managed on a weekly basis
and 606 on a monthly basis. The following two-point summary
of results is indicative of the suitability of the gamma distribution
on one hand and the irregularity of demand on the other. At the
same time it shows that the DSS covers comprehensively all the
different situations and provides solutions using a variety of tools,
which have been developed and presented in the preceding
sections.

Final step. Output
The program prints out the results in the form shown in Table 4,
which contains a sample output for seven disguised items. The
information printed for every SKU consists of
 Key input data: code number (the first letter denotes the
supplier), time unit of the demand data and analysis
(week or month), review period R and lead time L.
 Demand characteristics: the distribution that shows the
best fit to the demand data (d.fitted), the distribution that
was eventually used in the analysis (d.used), as well as
E(D), CV and P(D > 0).
 Results: base stock S, expected stock on hand E(OH), and
actual fill rate (FR). The reported FR is often somewhat
higher than the target fill rate b due to upward rounding
of S, as is the case for item B18X7 in Table 4 (target
b ¼ 0:95, actual FR = 0.98).
9. Results and conclusions
The decision support system (DSS), which has been developed
to assist RODA in managing their inventories, is currently in full
operation. It cooperates smoothly with the ERP system that the
company installed in 2007; the ERP system feeds the DSS with de-
mand information and the DSS returns the recommended base
stock values S. These values are updated every 6 months, using
the most current two-year demand data in a rolling-horizon fash-
ion. At the specific review instances, which are determined for each
supplier according to the respective review period, the ERP uses
the S values to derive the appropriate order quantities for every
supplier, taking into account the current inventory positions of
the various SKUs. When necessary, the recommended order quan-
tities are adjusted, mostly upwards, to meet the requirements of
minimum economic quantities per item or per supplier, so as to
take advantage of quantity discounts.
The first year of implementation of the combined DSS/ERP sys-
tem was 2008. At the end of that year the system was being used to
issue orders to all 23 regular suppliers. Each order consists on aver-
age of 9-10 line items; 30% of those correspond to planned orders
by RODA’s customers, which are inserted to the ERP outside and
independently of the DSS system, while the remaining 70% of the
line items concerns SKUs managed to stock through the DSS sys-
tem. At the time that this paper was being written the most recent
execution of the program concerned the re-evaluation of recom-
 The gamma distribution provided a good fit for 47% of the
1420 SKUs, including more than half of the SKUs managed
on a weekly basis, which have on average higher annual
demand. An almost equal percentage of SKUs had insuffi-
cient data to fit a distribution and most of them had less
than three non-zero demand periods, leading the program
to recommend S = 0 for 375 of the 1420 SKUs. The
remaining cases were handled mostly with the aid of
the package Poisson distribution, while the regular Pois-
son was found to be the most satisfactory choice in only
3 of the 1420 cases.
 The program computed base stock values using the math-
ematical expressions of Section 6 for 751 SKUs, while it
resorted to the simple heuristic rules of the form
S ¼ AlBRþL for 294 SKUs with insufficient demand data.
In 114 of the 751 cases of the former category, though,
the program issued a warning that it is not recommended
to use target fill rates higher than 0.8 because the demand
of those SKUs is highly erratic, despite the fact that the
hypothesized fit to the gamma distribution with a spike
at zero or to the package Poisson distribution is not
rejected at the 1% significance level.
According to the top management of RODA, the inventory man-
agement system described in this paper helped the company’s
commercial business unit first and foremost to rationalize its pro-
curement procedures and manage its stock in a more systematic,
objective and transparent way. Second, it contributed to an en-
hanced understanding of the distinct characteristics of the differ-
ent types of SKUs and to an explicit examination of the relative
importance of the various SKUs, which is translated to specific tar-
get service levels. A side benefit of this exercise was the identifica-
tion of obsolete SKUs and a streamlined catalogue of 10% fewer
items. The third and probably most concrete economic benefit of
the new system is that it has helped the company to reduce its
inventories without sacrificing the provided service, by identifying
and removing excess safety stock from many SKUs.
More specifically, RODA reports the following specific improve-
ments in 2008 compared to 2007, which are all attributed to the
new inventory management system:
(a) The average stock on hand has decreased by about 8% in
the one year since the beginning of the DSS/ERP imple-
mentation, without any noticeable deterioration of the

service level and despite a 3% growth in the sales volume
of the company’s commercial business unit during the
same time period.
(b) The transportation cost as a fraction of the total cost of
orders to RODA’s suppliers has been reduced from 7% in
2007 to 5.2% in 2008.
(c) The number of unplanned orders to the suppliers
decreased by 22%, from 18 orders per month on average
in 2007 to 14 in 2008.
(d) According to the customer satisfaction questionnaire,
which RODA sends every year to its customers asking
them to evaluate it along several dimensions on a scale
from 2 (worst) to +2 (best), the average rating of the
order response speed increased from 1.36 in 2007 to
1.70 in 2008.
In summary, points (a), (b), (c) and (d) above indicate collec-
tively that all involved parties are better off with the new system:
not only the company itself (a and b), but also its suppliers (c) and
customers (d).
Acknowledgements
The authors thank the management and staff of RODA s.a. for
their close collaboration throughout this project as well as for their
permission to publish real data.
Appendix A. Derivation of expected stock on hand in an (R, S)
system
Given that an order up-to-S is placed L time units prior to the
beginning of the cycle and is received just before the cycle begins
ðt1 Þ, the net stock at the beginning of the cycle ðNSB Þ is positive
whenever the demand xL during the lead time L is lower than S
and negative otherwise. In addition, since no other order is re-
ceived during the cycle of length R, the net stock at the end of
the cycle ðNSE Þ is positive whenever the demand xRþL during R+L
is lower than S and negative otherwise.
Case B of Fig. 2 occurs whenever NSB > 0ðxL < SÞ and
NSE < 0ðxRþL > SÞ, that is with probability
Z S Z S Z SxL
pA ¼ fRþL ðxRþL ÞdxRþL ¼
0 0 0 Poisson demand. Management Science 12 (5), 391–411.
Case B of Fig. 2 occurs whenever NSB > 0 ðxL < SÞ and
NSE < 0 ðxRþL > SÞ, that is with probability
Z S Z 1
pB ¼ fR ðxR ÞfL ðxL ÞdxR dxL :
0 SxL
Case C of Fig. 2 occurs when xL < S and consequently NSB < 0
and NSE < 0, that is with probability
Z 1
pC ¼ fL ðxL ÞdxL :
S Janssen, F., Heuts, R., de Kok, T., 1998. On the (R, s, Q) inventory model when
Inventory is available on hand during the entire cycle in case A
and part of the cycle in case B of Fig. 2. Its average magnitude de-
pends on the expected values of NSB and NSE in those cases. The
conditional expectations of NSB in cases A and B are denoted
EA ðNSB Þ and EB ðNSB Þ, respectively, and are given by
Z Z
S SxL
EA ðNSB Þ ¼ ðS  xL ÞfR ðxR ÞfL ðxL ÞdxR dxL =pA ;
0 0
Z S Z 1
EB ðNSB Þ ¼ ðS  xL ÞfR ðxR ÞfL ðxL ÞdxR dxL =pB :
0 SxL
G. Nenes et al. / European Journal of Operational Research 205 (2010) 313–324
fR ðxR ÞfL ðxL ÞdxR dxL :
323
The conditional expectations of NSE in cases A and B are denoted
EA ðNSE Þ and EB ðNSE Þ, respectively, and are given by
Z S
EA ðNSE Þ ¼ ðS  xRþL ÞfRþL ðxRþL ÞdxRþL =pA ; ðA:6Þ
0
Z S Z 1
EB ðNSE Þ ¼ ðS  xR  xL ÞfR ðxR ÞfL ðxL ÞdxR dxL =pB : ðA:7Þ
0 SxL
The expected stock on hand in case A, EA ðOHÞ, is given by
EA ðOHÞ ¼ ½EA ðNSB Þ þ EA ðNSE Þ=2 ðA:8Þ
because there is inventory available on hand during the entire cycle.
The expected stock on hand in case B, EB ðOHÞ, is given by
EB ðNSB Þ EB ðNSB Þ
EB ðOHÞ ¼  ðA:9Þ
2 EB ðNSB Þ  EB ðNSE Þ
because the average stock EB ðNSB Þ=2 is available only during the
part of the cycle with positive net stock. The expected value of this
portion is EB ðNSB Þ=½EB ðNSB Þ  EB ðNSE Þ.
Taking all the above into account, the expected stock on hand
can be calculated from EðOHÞ ¼ pA EA ðOHÞ þ pB EB ðOHÞ. Using (A.1)–
(A.9) and simplifying, we end up with expression (1) for E(OH).
References
Aronis, K.-P., Magou, I., Dekker, R., Tagaras, G., 2004. Inventory control of spare parts
using a Bayesian approach: a case study. European Journal of Operational
Research 154 (3), 730–739.
Bartezzaghi, E., Verganti, R., Zotteri, G., 1999. A simulation framework for
forecasting uncertain lumpy demand. International Journal of Production
Economics 59 (1–3), 499–510.
Burgin, T.A., Wild, A.R., 1967. Stock control – experience and usable theory.
Operational Research Quarterly 18, 35–52.
Cheung, K.L., 1996. On the (S  1, S) Inventory model under compound Poisson
demands and i.i.d. unit resupply times. Naval Research Logistics 43 (4), 563–
572.
Croston, J.D., 1972. Forecasting and stock control for intermittent demands.
Operational Research Quarterly 23 (3), 289–303.
Dolgui, A., Pashkevich, M., 2008. Extended beta-binomial model for demand
forecasting of multiple slow-moving inventory items. International Journal of
Systems Science 39 (7), 713–726.
Dunsmuir, W.T.M., Snyder, R.D., 1989. Control of inventories with intermittent
demand. European Journal of Operational Research 40 (1), 16–21.
Ehrhardt, R., 1979. The power approximation for computing (s, S) inventory policies.
Management Science 25 (8), 777–786.
ðA:1Þ Feeney, G.J., Sherbrooke, C.C., 1966. The (S  1, S) inventory policy under compound
Forsberg, R., 1995. Optimization of order-up-to-S policies for two level inventory
systems with compound Poisson demand. European Journal of Operational
Research 81 (1), 143–153.
Friend, J.K., 1960. Stock control with random opportunities for replenishment.
Operational Research Quarterly 11 (3), 130–136.
Gelders, L.F., van Looy, P.M., 1978. An inventory policy for slow and fast movers in a
ðA:2Þ petrochemical plant: a case study. Journal of the Operational Research Society
29 (9), 867–874.
Ghobbar, A.A., Friend, C.H., 2002. Sources of intermittent demand for aircraft spare
parts within airline operations. Journal of Air Transport Management 8 (4),
221–231.
Ghobbar, A.A., Friend, C.H., 2003. Evaluation of forecasting methods for intermittent
parts demand in the field of aviation: a predictive model. Computers and
ðA:3Þ Operations Research 30 (14), 2097–2114.
demand is modelled as a compound Bernoulli process. European Journal of
Operational Research 104 (3), 423–436.
Johnston, F.R., Boylan, J.E., 1996. Forecasting for items with intermittent demand.
Journal of the Operational Research Society 47 (1), 113–121.
Kukreja, A., Schmidt, C.P., 2005. A model for lumpy demand parts in a multi-location
inventory system with transshipments. Computers and Operations Research 32
(8), 2059–2075.
Porras, E., Dekker, R., 2008. An inventory control system for spare parts at a refinery:
an empirical comparison of different re-order point methods. European Journal
of Operational Research 184 (1), 101–132.
ðA:4Þ Rao, A.V., 1973. A comment on: forecasting and stock control for intermittent
demands. Operational Research Quarterly 24 (4), 639–640.
Regattieri, A., Gamberi, M., Gamberini, R., Manzini, R., 2005. Managing lumpy
ðA:5Þ demand for aircraft spare parts. Journal of Air Transport Management 11 (6),
426–431.
324 G. Nenes et al. / European Journal of Operational Research 205 (2010) 313–324
Segerstedt, A., 1994. Inventory control with variation in lead times, especially when
demand is intermittent. International Journal of Production Economics 35 (1-3),
365–372.
Silver, E.A., Pyke, D.F., Peterson, R., 1998. Inventory Management and Production
Planning and Scheduling. Wiley, New York.
Silver, E.A., Robb, D.J., 2008. Some insights regarding the optimal reorder period in
periodic review inventory systems. International Journal of Production
Economics 112 (1), 354–366.
Strijbosch, L.W.G., Heuts, R.M.J., Luijten, M.L.J., 2002. Cyclical packaging planning at
a pharmaceutical company. International Journal of Operations and Production
Management 22 (5), 549–564.
Syntetos, A.A., Boylan, J.E., 2005. The accuracy of intermittent demand estimates.
International Journal of Forecasting 21 (2), 303–314.
Syntetos, A.A., Boylan, J.E., 2006. On the stock control performance of intermittent
demand estimators. International Journal of Production Economics 103 (1), 36–
47.
Thompstone, R.M., Silver, E.A., 1975. A coordinated inventory control system for
compound Poisson demand and zero lead time. International Journal of
Production Research 13 (6), 581–602.
Teunter, R., Sani, B., 2008. Calculating order-up-to levels for products with
intermittent demand. International Journal of Production Economics.
doi:10.1016/j.ijpe.2008.08.012.
Vereecke, A., Verstraeten, P., 1994. An inventory management model for an
inventory consisting of lumpy items, slow movers and fast movers.
International Journal of Production Economics 35 (1-3), 379–389.
Watson, R.B., 1987. The effects of demand-forecast fluctuations on customer service
and inventory cost when demand is lumpy. Journal of the Operational Research
Society 38 (1), 75–82.
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 (3), 375–387.
Willemain, T.R., Smart, C.N., Shockor, J.H., DeSautels, P.A., 1994. Forecasting
intermittent demand in manufacturing: a comparative evaluation of Croston’s
method. International Journal of Forecasting 10 (4), 529–538.
Williams, T.M., 1984. Stock control with sporadic and slow-moving demand. Journal
of the Operational Research Society 35 (10), 939–948.
Zhao, Y., 2008. Evaluation and optimization of installation base-stock policies in
supply chains with compound Poisson demand. Operations Research 56 (2),
437–452.
Zipkin, P., 1986. Inventory service-level measures: convexity and approximation.
Management Science 32 (8), 975–981.