Inventory management

Citation for published version (APA):

de Kok, T. (2018). Inventory management: Modeling real-life supply chains and empirical validity. Foundations
and Trends in Technology, Information and Operations Management, 11(4), 343-437.
https://doi.org/10.1561/0200000057

Document license:
Unspecified

DOI:
10.1561/0200000057

Document status and date:

Published: 23/04/2018

Document Version:
Accepted manuscript including changes made at the peer-review stage

Please check the document version of this publication:

• A submitted manuscript is the version of the article upon submission and before peer-review. There can be
important differences between the submitted version and the official published version of record. People
interested in the research are advised to contact the author for the final version of the publication, or visit the
DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page
numbers.
Link to publication

General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

• Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.

If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please
follow below link for the End User Agreement:
www.tue.nl/taverne

Take down policy

If you believe that this document breaches copyright please contact us at:
[email protected]
providing details and we will investigate your claim.

Download date: 16. Jan. 2025

Inventory Management: Modelling Real-life
Supply Chains and Empirical Validity

Ton de Kok

Ton de Kok
School of Industrial Engineering, Eindhoven University of Technology, e-mail: [email protected]

1
Contents

Inventory Management: Modelling Real-life Supply Chains and

Empirical Validity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Ton de Kok
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Modelling inventory systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1 A brief history of inventory management research . . . . . . 8
2.2 Empirical validity of inventory models . . . . . . . . . . . . . . . . 9
2.3 The practice of inventory management: human
intervention and correlations . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Intervention frequency and inventory system performance 11
2.5 Modelling time between order release and order receipt . . 12
2.6 Conclusion on modelling inventory system . . . . . . . . . . . . 14
3 Single-item single-echelon inventory models . . . . . . . . . . . . . . . . . . 14
3.1 Deterministic demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Stochastic demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4 Uncapacitated multi-item multi-echelon models . . . . . . . . . . . . . . . . 27
4.1 Material availability and stochastic lead times . . . . . . . . . . 28
4.2 The example supply chain . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.3 Modelling multi-echelon inventory systems . . . . . . . . . . . . 30
4.4 Feasibility of order release quantities . . . . . . . . . . . . . . . . . 32
4.5 Synchronization and allocation . . . . . . . . . . . . . . . . . . . . . . 35
4.6 Decision node structure for the case example . . . . . . . . . . 42
4.7 Control policies for divergent MIME systems . . . . . . . . . . 47
4.8 Generalized Newsvendor equations for divergent
MIME systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.9 Performance of SBS policies . . . . . . . . . . . . . . . . . . . . . . . . 51
4.10 Empirical validity of SBS policies . . . . . . . . . . . . . . . . . . . 54
4.11 Positioning inventory in the supply chain . . . . . . . . . . . . . . 57
5 Capacitated inventory systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.1 Feasibility of order release quantities . . . . . . . . . . . . . . . . . 60
5.2 Comparison of rolling scheduling concepts . . . . . . . . . . . . 62

3
4 Contents

5.3 Optimal policies for serial MIME systems . . . . . . . . . . . . . 62

5.4 Implicit modelling of finite capacity . . . . . . . . . . . . . . . . . . 63
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
7 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

1 Introduction

Inventory management has been a core topic of Operations Research since the
1950s. Inventory can be seen as a means to create efficiency in production and dis-
tribution: it enables scale by allowing to accumulate demand until a batch quantity
can be released that can be produced and shipped efficiently. This role of inventory
is of great importance in process industries, where set-up times are considerable.
Inventory can be seen as a means to ensure sufficient customer service: as demand
is unpredictable we must hold inventory just-in-case, to protect against unexpected
surges in demand. This role of inventory is of great importance in retail, as we expect
a product to be available off-the-shelf or at our doorstep within 24 hours.
Inventory can also be seen as a symptom of bad management, as waste of capi-
tal. Reduction of inventory capital has been high on the priority lists of CEO’s over
the last four decades. Early 1980’s the Just In Time (JIT) philosophy proclaimed
zero inventory as the key objective to ensure continuous improvement of processes,
leading to less process variability, shorter processing time, smaller production and
transportation batches, and higher product yield. In many businesses inventory is
a forbidden word. Euphemisms for inventory were introduced, such as buffers, and
supermarkets. Despite the continuous efforts to reduce process durations and volatil-
ity, zero inventory will remain a mirage, as fundamental uncertainty in demand and
supply cannot be eliminated and trading-off efficiency, quality, customer service and
cost of inventory capital inevitably yields the need for inventory at various places in
global and local supply chains, acting as the lubricant.
The trade-offs to be made have been studied extensively in the inventory manage-
ment literature. This has lead to optimal inventory control policies for various supply
chain structures under various cost assumptions. Clearly, most results are known for
the simplest inventory management situation, i.e. a single product at a single loca-
tion. But both the qualitative and quantitative understanding of this simple inventory
management situation is a building block for understanding inventory management
in practice, where we have to deal with multiple items at multiple locations.
Thus, inventory control policies are implemented in every ERP (Enterprise Re-
source Planning) system, e.g. SAP and Oracle, and used at almost every company.
ERP systems are the transaction backbone systems of enterprises in which prod-
uct and process data are stored, and each customer order, production order, and
purchase order is tracked and traced. Over the course of a few decades ERP sys-
tems have been enriched with planning and control modules that support inventory
management, production management, and sales. Despite the availability and use
Contents 5

of inventory control policies in ERP systems, we observe that most of the control-
policy-based replenishment proposals are overwritten by manual decisions. Indeed,
being an inventory manager or planner, you want to manage and plan, and you can
do better than the inventory management system. Unfortunately, it is shown again
and again that proper use of inventory management systems yields higher service
and lower costs at the same time. We observe that inventory managers have diffi-
culties with the interpretation of unexpected events regarding demand and supply,
i.e. distinguishing noise from signal. At the same time we observe that inventory
managers have access to relevant information that an ERP system’s inventory con-
trol module cannot exploit. This calls for the design of an inventory management
approach that combines the strength of mathematically rigorously determined in-
ventory control policies and tacit knowledge of human decision makers. This paper
is motivated by these observations and builds on 32 years of working in (8 years)
and with (the next 24 years) industry, applying and implementing inventory control
models.
It is our intention to write a different overview of inventory models, from single
item single echelon models to multi-item multi-echelon models, then is mostly pro-
vided in text books on Operations Management(e.g. Nahmias & Olsen (2015) and
Silver et al. (2016)). We hope that this paper provides complementary knowledge.
Instead of starting with inventory models that are tractable from a mathematical
point of view, we start from the inventory management problem and the modelling
challenges to be faced.
The first section of this paper is devoted to modelling inventory systems, such
that these models are empirically valid by proper calibration. Inventory models are
abstractions that cannot capture all possible actions to balance supply and demand.
But with proper measurement of inventory management performance we can set
the parameters in such a way that the customer service is consistently at the right
level. We hypothesize that it is better to use mathematically tractable models and
appropriately chosen performance measures, then to identify all possible actions
under specific circumstances and model these explicitly. We found that many spe-
cific actions are focussed on preventing stockouts. Typically, such actions either
postpone customer demand or expedite production orders released earlier. Herewith
we create correlation between occurrences of high demands and arrivals of produc-
tion orders that satisfy them. Ignoring this correlation yields considerable underes-
timates of customer service, while modelling this correlation is mostly mathemati-
cally intractable. Thus we propose to measure performance before specific actions
are taken, which yields the notion of Intervention Independent Performance (IIP)
indicators. Clearly, a company must measure the effectiveness of the specific ac-
tions taken as well, which yields the notion of Intervention Dependent Performance
(IDP) actions. Applying IIP indicators in combination with inventory models in re-
search projects provided an empirical basis for the validity of this approach: in both
single-item single-echelon (SISE) situations and multi-item multi-echelon inventory
systems we could explain the quantitative relationship between capital invested in
item inventories and end-item customer service. One should not underestimate the
importance of this finding: it provides a scientific basis for the use of inventory mod-
6 Contents

els as studied in OR literature. Here we take the position that mathematical models
and their analysis are not science without empirical data supporting the causalities
embodied by the model.
The second section discusses SISE models. We show that under linear holding
and penalty costs the Newsvendor equation holds for virtually any sensible control
policy. The Newsvendor equation states that the non-stockout probability at an ar-
bitrary point in time equals the quotient of penalty cost rate and the sum of holding
cost rate and penalty cost rate. We show that inventory management performance is
primarily determined by average inventory and order frequency. In our view there
should be more emphasis in inventory management education on average inventory
levels instead of safety stocks. After all, we pay for the capital tied up in average
inventories, not in safety stocks. As capital is tied up in inventory, it is relevant to
consider trade-offs from a Return On Investment point of view. We discuss the im-
pact of the change from cost minimization to ROI maximization using the Economic
Order Quantity model. We discuss the prerequisites for empirical validity of the ba-
sic inventory models. One lesson should stand out here: mathematical analysis must
be rigorous. Otherwise it is likely that the resulting control policies do not make any
sense to inventory planners, and they are right in that case.
The third section extensively discusses multi-item multi-echelon inventory sys-
tems. This discussion is not aiming at a complete overview of the state-of-the-art
on multi-echelon inventory system research. Having worked on the subject for over
25 years, we conclude that the emphasis in the scientific literature has primarily
been on optimal policies under specific assumptions on the structure of multi-item
multi-echelon systems, such as serial, divergent or convergent, (cf. Axsäter (2003)
and Song & Zipkin (2003)) and much less on the underlying complexity of general
multi-item multi-echelon (MIME) systems. There are no serial systems in prac-
tice! At best they are divergent (i.e. each item has a single upstream predecessor, or
child) in the form of retail and spare parts distribution networks. Convergent MIME
systems, i.e. systems in which each item has at most one parent, are rare, as most
companies sell more than one product. In literature, convergent MIME systems are
also referred to as (pure) assembly systems. So most of the time supply chains are
networks with both embedded divergence and convergence (i.e. an item may have
multiple children upstream and multiple parents downstream. Under uncertainty you
are continually confronted with the dilemma to allocate item availability among par-
ent items, i.e. the items that use the item under consideration. Allocating less to a
particular parent item implies that less is needed of other child items used by this
parent item, whereby these child items can be used for other parent items, but then
we need other items as well, etc. We assume that orders released to the shopfloor can
be executed with 100% due date reliability, provided material (and resource) con-
straints are taken into account. This implies that we model general MIME systems
with constant flow times, i.e. constant times between order release and order receipt
in inventory. In order to create a benchmark for control policies for general MIME
systems, we formulate necessary conditions for a control policy to yield feasible
solutions. Herewith we bridge the gap between mathematical programming formu-
lations of supply chain planning problems that concern the problem to be solved
Contents 7

today, and the stochastic dynamic programming formulations that focus on control
policy structures that generate optimal policies, and resulting solutions, over a rele-
vant period of time.
The most-often used planning logic to plan and manage multi-item multi-echelon
inventory systems in practice is called Material Requirements Planning, which is
abbreviated as MRP I. The main principles of MRP I logic are lead-time offsetting
and dependent demand. Starting from the constraint to maintain a safety stock at
the end of each future period, and known future (gross) requirements for an item,
and outstanding orders, inventory balance equations are used to determine the re-
plenishment quantities in future period. By offsetting the replenishment quantities
by the item lead time we obtain planned order quantities. These planned order quan-
tities are translated into so-called dependent demand for child items by multiplying
the order quantities by the number of child items needed to make one item. By
proper administration we can determine the dependent demand for each item and
derive the planned order for each item. For further details on the logic we refer to
subsection 4.4. Initially Material Requirements Planning was abbreviated as MRP,
but in the 1980s the MRP logic was embedded in an overall framework for plan-
ning and control called Manufacturing Resource Planning, which, having the same
three-letter-abbreviation, was denoted as MRP II (cf. Vollmann et al. (2005)). MRP
I was introduced as a ”killer app” for IBM mainframes in the early 1960s, and pro-
moted by the American Production and Inventory Control Society (APICS) from
1970 onwards. For a historic perspective on MRP I we refer to Wilson (2016). We
find that MRP I logic does not pass the test of adhering to material availability con-
straints. This finding cannot be emphasized often enough, as it explains symptoms
like nervousness and expediting. On my return to academia early 1990s I set myself
the research objective to determine safety stocks in MRP I. Pursuing this objective,
I found that my quest would be in vain, because the MRP I logic is not mathemati-
cally sound. MRP I logic turned out to be a logic that generates requirements, but it
is not a logic for planning. Planning involves the balancing of demand and supply,
knowing that you must take decisions on supply before demand is known. That is
why in general MIME systems there is a continual misalignment between demand
and supply that is resolved by keeping inventory. But inventory does not always
resolve the misalignment, and that is where scarce child item material availability
must be allocated among multiple parent items, with the consequences sketched
above: a problem mess, a Gordian knot. The concept of Synchronized Base Stock
(SBS) policies for operational control of general multi-item multi-echelon inven-
tory systems, is cutting this Gordian knot at the expense of suboptimality (though
SBS policies are optimal for divergent systems, and convergent systems). The SBS
concept generates a deep insight into the natural decision hierarchy embedded in
any general multi-item multi-echelon system. In-depth case studies in the context
of MSc thesis projects at companies indicate that the assumption of SBS policies
yields empirically valid results, even though none of these companies used SBS
policies. The only explanation for this result is that also in multi-item multi-echelon
inventory systems inventory performance is driven by average inventories and order
frequencies.
8 Contents

The fourth section briefly discusses the additional issues that come with taking
into account resource constraints. While for single echelon systems finite capacity is
(relatively) easy to do this, this is not the case for multi-echelon systems. I consider
the results for serial systems in Janakiraman & Muckstadt (2009) as a milestone in
the analysis of capacitated multi-item multi-echelon systems, and at the same time
as a clear indication of the challenges ahead of us when trying to tackle this problem
for general structures.
Inventory management is a challenging research subject due to its structural com-
plexity, represented by general networks of interacting stockpoints, and the com-
plexity induced by demand and supply uncertainty. The curses of dimensionality
prohibit the calculation of optimal policies. I hope that this fact is a reason to pursue
more research with great practical relevance. Admittedly, when allowing yourself to
write down that something on the left hand side of an ”equation” is approximately
equal to something on the right hand side, you may be overwhelmed by the possi-
ble alternative routes that can be taken towards policies and algorithms. But at the
end of the day, applied science should be about reality and reality happens to be
complex.

2 Modelling inventory systems

In this section we focus on modelling of inventory systems as opposed to analysis

of inventory models. We first provide a brief overview of the inventory management
research since 1888. In subsection 2.2 we explain our view on empirical validity
of inventory models (and stochastic models in general). We discuss how human
interventions have a great impact on the empirical validity of inventory model in
subsection 2.3. In subsection 2.4 we propose to develop performance measures that
cannot be impacted by human interventions, so that they can be used to validate and
calibrate inventory models. In subsection 2.5 we discuss the definition and mea-
surement of time between order release and order completion and receipt, as proper
modelling of this time interval is quintessential for the validity of inventory models.

2.1 A brief history of inventory management research

Quantitative models for Inventory Management has been an important subject in the
context of Operations Management for about 130 years. The first publication on the
subject is by Edgeworth (1888), who studies the Newsvendor problem in the con-
text of a cash balance problem. The derivation of the Economic Order Quantity has
been claimed by several different authors from different language backgrounds, but
Harris (1913) seems to be the first source. Interesting to note here that the most ba-
sic stochastic inventory system has been studied before the most basic deterministic
inventory system. The start of Inventory Management research as we know it today
Contents 9

can be attributed to Whitin (1953). The analysis of SISE systems is extensive and
seems rather complete. It is a standard subject in IE, OM, and OR curricula, and
algorithms to determine control policies have been embedded in inventory man-
agement systems, either as standalone applications or as modules of ERP systems.
The analysis of multi-item multi-echelon systems has progressed substantially since
1990, but stochastic demand and multiple items to be managed in some coordinated
and cost-effective manner implies that we face the curses of dimensionality. There
is no hope to find optimal control policies for real-world multi-item multi-echelon
systems. But the progress made on the analysis of these complex inventory sys-
tems is such that we conjecture that state-of-the-art inventory systems research can
support the quantitative analysis of real-world supply chains under demand and sup-
ply uncertainty. The emergence of supply chain optimization software over the last
two decades, and the extensive documentation of its successful application in lit-
erature, support this conjecture. This supply chain optimization software is seen as
an example of so-called Advanced Planning and Scheduling (APS) systems. These
APS systems are using optimization methods and heuristics to solve various kinds
of deterministic planning problems. APS systems are used to solve facility loca-
tion problems, aggregate planning problems with a horizon of 2-5 years, mid-term
planning problems with a horizon of 1-2 years, but also shopfloor scheduling prob-
lems with a horizon of a few weeks or days, even. Thus the outcome of most APS
systems is a plan. Inventory research based supply chain optimization software typi-
cally produces inventory control parameters, explicitly taking randomness in supply
and demand into account. These inventory control parameters are used in the APS
planning software to guide decision making. Inventory control parameters are safety
stocks, safety times, lot sizes, and review periods. For an extensive overview on APS
systems and their application, we refer to Kilger et al. (2015).

2.2 Empirical validity of inventory models

Empirical validity is a central concept in science. We adopt here the concept as used
in Physics, as inventory models are models describing causality between inputs and
outputs quantitatively, similar to Bohr’s model of the atom describes emission of
quanta of energy as a consequence of energy added, and Newton’s model of gravity
describes the speed of an object falling towards the earth. In Physics it is gener-
ally assumed that measurement errors are normally distributed, so that a confidence
interval around the point estimate of the variable under consideration can be deter-
mined. A model is empirically valid if the value predicted by the model is within
the confidence interval. Experiments are repeated as many times as needed under
the same conditions to produce a supposedly identical outcome.
Validation of inventory models cannot be done this way. Firstly, inventory man-
agement takes place in reality, not in a lab. Conditions under which the control
policies derived from the inventory model are executed cannot be controlled. Sec-
ondly, as demand and supply are stochastic by nature, a single experiment concerns
10 Contents

a item’s inventory process over a period of time, typically at least half a year. Over
this period data are collected from which probability distributions are determined to
describe the demand and supply processes. As a probability distribution is funda-
mentally a concept relating to repeating an experiment infinitely many times under
the same conditions, a model error is inevitable. The probability distributions de-
rived have nothing to do with measurement errors, they are elements of the model
itself. Thirdly, we cannot repeat experiments, as product life cycles are finite.
In our empirical studies we adopted the following approach: we collected data
over a some period of time for a (large) number of different items. We assumed
stationary stochastic processes and derived from the data estimates for the mean
and standard deviation of the random variables involved. In the context of inventory
management models, these random variables are demand per time unit for the sit-
uation of periodic review of inventory, order interarrival times and order sizes for
the situation of continuous review of inventory, and replenishment order lead time.
Furthermore we computed average inventories and average order sizes for each item
over the same period of time. Under some control policy, we determine the policy
parameters that yield the average inventory measured. Given those policy parame-
ters we compute the customer service according to the model and compare against
the actual customer service over the data collection period. So we execute the same
experiment for different items.
This approach can be compared with throwing dices 52 times, say, and computing
the probability that the resulting outcome is at most 5. According to probability
theory the outcome should be 5/6, but in reality this is not the case. If we execute
the experiment with a large number of different dices, we may hope that our model
provides an accurate aggregate outcome. When considering multiple different items
we weighted the outcomes with the financial turnover of each item. This may be
considered adding apples and oranges, but we want to assess the validity of a generic
model that can be applied to many different inventory control situations.
When we claim empirical validity of models in the sequel of this paper, it should
be understood in the way described above. Clearly, more research is needed to de-
velop a rigorous methodology to validate stochastic models.

2.3 The practice of inventory management: human intervention

and correlations

In this section we provide an overview of the state-of-the-art of quantitative inven-

tory management research from a professional perspective. The professional per-
spective is that of an inventory manager who needs to create a setting in which in-
ventory planners, supply planners, purchasers and expediters can operate effectively.
Supply must be balanced with demand to the extent possible and the joint efforts of
manager, planners, and expediters represent a giant array of possible means to that,
even when demand for products comes unexpectedly. Our starting point is that we
cannot model this array of options to respond to unexpected demand (and unexpect-
Contents 11

edly delayed supplies). We need to develop models that support the professionals
involved in inventory management to exploit their array of options. In this intro-
duction we provide arguments that support the application of the inventory theory
developed over more than a century.
When applying inventory models in practice we often found that application of
the textbook formulas for safety stocks and customer service provided in e.g. Fog-
arty & Hoffmann (1983) shows the results are empirically invalid. In many cases
this could be explained by the invalidity of the assumptions underlying these text-
book formulas and the invalidity could be resolved by relaxing assumptions and
rigorous mathematics (cf. De Kok (1991a) and Section 2). However, in a substantial
number of cases the invalidity could not be resolved. We found that the common
denominator of these cases was human intervention: planners schedule-in orders
when needed, expediters expedite these orders and reallocate materials to produc-
tion orders to prevent belt stops. In general, human interventions create correlation
between supply and demand: if demand is high, lead times are shorter. If demand
is high, demand is aligned with supply by negotiating a later delivery date than re-
quested. We stated above that we cannot model all response options. On top of that,
taking into account correlation between stochastic demand and stochastic supply is
mathematically inhibitive in most cases.

2.4 Intervention frequency and inventory system performance

Based on our extensive involvement in development of models for inventory systems

in practice we conclude that, though we cannot model all responses to uncertainty,
and we cannot model correlation between demand and supply, we can model the
triggers for non-modelled responses to uncertainty and effect of responses to un-
certainty. The main trigger for non-modelled responses to uncertainty are projected
shortages within the supply lead time, where projected refers to both as derived
from an algorithm, and to tacit knowledge of the professional. The projected short-
ages can be both immediate shortages and future shortages within the lead time.
Assuming we know the events that trigger non-modelled response options, we can
link these trigger events to events in the inventory model. In particular we may
consider stockout events and customer backlog events as trigger events. Assuming
that before such events occur, the inventory system is controlled according to the
control policy modelled, we can use the frequency of trigger events in reality to
the frequency of stockouts or the probability a customer must wait, respectively, in
our model. Assuming that the trigger frequency does not change in the future, we
can use this frequency as a performance target in our inventory model to derive the
parameters of the inventory control policy.
As each trigger event enacts a response from the planner, expediter, or purchaser,
we must link the response enacted to its effect, which should be prevention of a
shortage or reduction of customer waiting time. In the case situation that has been
the source of the above reasoning, the schedule-in actions resulted into 100% cus-
12 Contents

tomer service with one month of average inventory. No standard inventory model
assuming normally or gamma distributed demand per period could explain this.
However, by assuming that each schedule-in actions prevented a stockout, we could
determine control policies from the number of stockouts per unit time. The empir-
ical validation of the model supported the modelling approach (cf. De Kok (1998)
and de Kok (2017)). In this case the effect of the response actions was unambigu-
ous: 100% customer service. In other cases we may find an increase of customer
service from 80% according to the inventory model based on historical data about
the demand process, supply process and inventory process, to 98% in reality. In that
case we need to make additional modelling assumptions to explain the 18% increase
in service. Suppose we use the non-stockout probability at an arbitrary point in time
as service measure. Then it should be that without exploiting the non-modelled re-
sponse options, we would have found 80% service, while 90% of the responses was
effective, i.e prevented a stockout. Assuming that this 90% measures the capability
of the organization to prevent stockouts that cannot be prevented by the routine in-
ventory management rules, and assuming that this capability is the same in the (near)
future, we can use this capability estimate to derive the customer service target for
the inventory model from the customer service target in reality.
The critical step in this approach is the correct determination of the response
trigger frequency. Expediting an order may result in both expedition and postpone-
ment of other orders. Just determining which orders are produced or received earlier
than planned from an ERP system is not enough, as this may seriously overstate the
number of responses. But in most cases it is feasible to develop system support for
correct identification of non-modelled responses.
We already mentioned that the response trigger events can be related to events
in our inventory model. This leads to the concepts of intervention-independent and
intervention-dependent performance. The intervention-independent performance can
be predicted by mathematical analysis of inventory models, while this is impossi-
ble for the intervention-dependent performance without additional modelling as-
sumptions as discussed above. Fortunately, we only need to be able to measure the
actual intervention-independent performance, e.g. fill rate or non stockout probabil-
ity, to validate and calibrate our inventory models, so that they can be effectively
used to support inventory management professionals. For a detailed discussion of
intervention-independent and intervention-dependent performance indicators and
modelling manufacturing (and inventory) systems we refer to de Kok (2017). In
the remainder of this paper we assume that the performance targets to be achieved
relate to intervention-independent performance measures.

2.5 Modelling time between order release and order receipt

Before we conclude this discussion on modelling inventory systems and using data
from information systems to effectively validate and calibrate inventory models, we
would like to discuss modelling lead times. Over the course of time we find that
Contents 13

the concept lead time as defined in the inventory theory literature differs from the
concept lead time in the professional literature, and in particular in the APICS lit-
erature. In the latter body of knowledge lead time is a norm for the flow time as
used in planning systems, such as MRP II systems. The flow time of an order is the
time that elapses between release of an order and completion of the order. In the
scientific inventory theory literature the notion of lead time is used to describe the
time elapsed between release and completion of an order. Thus the lead time in the
scientific body of knowledge on inventory management is the same as the flow time
in the professional body of knowledge on inventory management. In this text we
use lead time as used in the scientific body of knowledge to denote the time elapsed
between order release and order completion (or receipt). In practice inventory man-
agement is not just ordering at the right moment in time the right quantity, but also
providing an outlook to the rest of the organization on future material availability.
Such an outlook can be derived if we make an assumption on the time that elapses
between release and completion of future orders. Thus a norm is introduced for the
lead time of future orders, which is denoted as nominal lead time (cf. Graves &
Willems (2003)). The nominal lead time can be derived as a sample mean of passed
lead times, or as a sample percentile, to include protection against uncertainty in
lead times.
In Zipkin (2000) exogenous and endogenous lead times are discussed. Exoge-
nous lead times are modelled as random variables independent of the other stochas-
tic processes. This is the most common way of modelling lead times in inventory
theory. Endogenous lead times are lead times that result from the interaction be-
tween stochastic demand and finite resources. These are typically discussed in the
context of queueing theory. Undoubtedly, lead times in real-life are endogenous. But
we should be aware that in most cases the lead time of an order is not determined
by its own workload, alone, but by the workload of many other orders. Queueing
models typically apply to situations with one or multiple resources and multiple
items that must be produced. From the perspective of a single item the lead time
is exogenous. Subsequent lead times may be dependent, as in most cases orders do
not overtake, but the lead time is not dependent on the individual item order size
or order interarrival time. So we aggregate item orders into one or a few job types
and apply queueing system analysis to determine an estimate of the lead time. This
is similar to aggregating data about item order lead times into an empirical distri-
bution of order lead times. When applying this approach in practice, we found that
queueing system models provide a good estimate of the average lead time, but the
standard deviation of order lead times is strongly overstated. The former could be
explained by the work conservation law in queueing theory (cf. Kleinrock (1965)),
the latter could be explained by similar arguments as above: shopfloor scheduling
professionals constantly keep track of the progress of production orders, and in-
tervene if the progress stalls for some reason, e.g. by assigning a high priority to
delayed orders. We should be aware that in queueing systems lead times just happen
as a consequence of the interaction between resource needs over time and resource
availability over time. In reality each production order has a due date and this due
date allows for feedback mechanisms. In reality not only can we reassign priorities
14 Contents

to orders, we can also reallocate resources to orders without affecting (too much)
the due date adherence of other orders.
As stated above, in reality planning systems use nominal lead times to enable
insight into the future evolution of inventory, and resource use. This insight is only
of use if the nominal lead times are realized. Therefore the nominal lead time de-
termines the due date of an order. And the due date provides a target that is used
to ensure the order is completed in time. At the shopfloor equifinality is the name
of the game: exploit all possible response options to ensure the order is finished
in time. It may be that the nominal lead time is set as the sum of the average lead
time as measured in the recent past (or according to a queueing system model) and
a safety time that provides the slack to meet the due date without excessive costs.
When modelling real-world supply chains with multi-item multi-echelon inventory
systems and constant lead times taken from ERP systems, we found that this ap-
proach yields empirically valid results. In fact, this empirical validity fostered the
above reasoning.

2.6 Conclusion on modelling inventory system

In summary, when modelling real-life systems, we should be aware that in reality

inventory management and shopfloor management professionals have an array of
response options to ensure timely delivery of customers and timely completion of
production orders, respectively. Trying to model these options is a dead-end due
to mathematical intractability, and from a conceptual point of view a fundamen-
tally wrong approach. By identifying the response trigger events and linking them
to events in the mathematical model, and defining and measuring the appropriate
intervention-independent performance measures we can establish a correct link be-
tween model and reality. The above arguments and their empirical foundation lead
to the conclusion that the inventory models developed over more than a century of
inventory management research can be used in practice. In the next sections we dis-
cuss the results from inventory research that we think are most important for appli-
cation in practice. The results selected and presented are based on highly subjective
decisions and should not be seen as a definitive selection.

3 Single-item single-echelon inventory models

In this section we discuss the basic SISE inventory models. We first discuss the sit-
uation in which we assume that demand per unit time is constant and known. This
is the situation discussed in Harris (1913). Instead of discussing the single-item
model, we discuss a multi-item single-echelon situation, allowing us to introduce in
a natural way another perspective on the problem than that of cost minimization: we
assume the firm wants to maximize its return on investment. This leads to some new
Contents 15

insights that were first published by Trietsch (1995), but in our view did not receive
sufficient attention. We also discuss the relation between EOQ and ABC classifica-
tion. Next we discuss the Newsvendor equation and how it unifies optimization of
SISE systems. We further show the robustness of stochastic inventory model perfor-
mance under given average inventory level and ordering frequency. This prepares
for the next section on multi-item multi-echelon systems.

3.1 Deterministic demand

In this subsection we consider the situation where a firm manages the inventory
of N items under the assumption of constant item demand rate. We assume linear
holding costs and fixed ordering costs for each item. We discuss the relationship
between the optimal order quantities and the so-called ABC classification, which
distinguishes between important and non-important items. We compare the optimal
order quantities under the objective of cost minimization with the optimal order
quantities under Return-On-Investment maximization.
Let us define

Table 1 Inventory system definitions, constant demand rate

Variable Definition
r yearly interest rate
Di yearly demand for item i, i=1,...,N
pi sales price per item i, i=1,...,N
vi cost price per item i, i=1,...,N
Ai fixed ordering cost of item i, i=1,...,N
F fixed assets
C sum of holding and ordering costs
R return on investment in fixed assets and inventory assets
QiC optimal order quantity of item i under cost minimization, i=1,...,N
fiC optimal order frequency of item i under cost minimization, i=1,...,N
Qi R optimal order quantity of item i under ROI maximization, i=1,...,N

3.1.1 EOQ and ABC classification

Following Harris (1913) we find that the optimal order quantity under minimization
of the sum of ordering and holding (capital) costs is given by

r
C 2Ai Di
Qi = , i = 1, ..., N (1)
vi r
16 Contents

N p
C=∑ 2Ai vi rDi (2)
i=1

Though the assumption of constant demand is very restrictive, many authors have
found that when extending the problem to stochastic demand and minimizing the
sum of ordering costs, holding cost and penalty costs, the expression for QiC yields
close to optimal results (cf. Silver et al. (1998)). It is also worth mentioning that it
follows from the findings in Daganzo (2005) that the expression for C is an accurate
approximation of the minimal yearly costs for the case of deterministic dynamic
item demand, provided the volatility is not too high.
Another relevant observation emerges when we consider the optimal ordering
frequency,

√
r
Di r
f iC = C
= Di v i , i = 1, ..., N (3)
Qi 2Ai

From equation (3) we observer that the optimal order frequency grows with the
square route of the item turnover. The higher the item order frequency, the more
often we must pay attention to the item, as it typically runs out of stock before arrival
of an order. Then it follows that the higher the item turnover, the more attention we
should pay to the item. This is fully in line with the so-called ABC classification,
where A-item are items with high turnover and C-items are items with low turnover.
The ABC classification assumes that we should pay most attention to A items. Thus
we see that the professional perspective behind ABC classification can be formally
supported by the Economic Order Quantity results of Harris (1913).

3.1.2 Cost minimization versus Return On Investment maximization

As mentioned above Trietsch (1995) takes a different perspective: maximization of

ROI. In that case we can derive the following objective function,

Qi vi r Ai Di
∑Ni=1 ((pi − vi )Di − 2 − Qi ) − Fr
R= (4)
F + ∑Ni=1 Q2i vi

By taking partial derivatives we can solve for the optimal QRi . It turns out to be
most convenient for solving the set of equations to introduce a constant λ ∗ , which
must satisfy the following set of equations,

s
2Ai Di
QRi = , i = 1, ..., N (5)
vi (λ + r)
Contents 17

QRi vi r
∑Ni=1 ((pi − vi )Di − 2 − AQi DR i ) − Fr
i
λ= (6)
QR v
F + ∑Ni=1 i2 i

The solution for the optimal λ is found by iteratively solving equations (5) and
(6) as a fixed point problem. The convergence is very fast, typically within 6 iter-
ations. Note that it follows from equations (4) and (6) that λ is the ROI itself! It
follows from equation (5) that if ROI is positive, then the economic order quantity
QCi is greater than the order quantity under ROI maximization QRi . It also follows
QC
from equation (5) that QiR is the same for all i.
i
In Trietsch (1995) a remarkable finding is reported that also holds for the for-
mulation of our ROI maximization problem that differs slightly from the one for-
mulated in his paper: if the relative annual demand quantities are fixed, then the
optimal QCi does not depend on the total annual demand (this follows directly from
the expressions for the partial derivatives, but is not obvious from equations (5) and
(6)). Thus the behaviour of the optimal lot sizes under cost minimization is com-
pletely different from that under ROI maximization. In the former case the lot size
grows as a square root of annual demand, while in the latter case it is independent
of the annual demand. This implies that under ROI maximization the frequency of
ordering increases linear in total annual demand, while under cost minimization the
frequency of ordering grows with a square root of annual demand. These statements
also hold when considering order frequency as a function of annual turnover. The
substantial difference between optimal lot sizes under cost minimization and ROI
maximization is illustrated in figure 1 and 2. We define the optimal ROI as R∗ and
R(EOQ) as the ROI when using QC for each item. The figures show the optimal
QC ∗
Return on Investment R∗ (right axis), ∆Q := QiR , and ∆ROI := R(EOQ)
R
as a function
i
of the fixed assets F.
The above findings are not widely reported in inventory management textbooks,
while in our view they are important and provide a broader perspective on the lot
sizing problem that is relevant to both students and professionals.

3.2 Stochastic demand

In this subsection we consider SISE models under stochastic demand. We assume

that the lot size can be derived from the EOQ model. Thereby we focus our discus-
sion on the control parameter that determines the amount of slack needed to cope
with demand and supply uncertainty, so that customer requirements can be satisfied
at minimal cost or at a sufficient level of customer service. Let us introduce some
notation.
18 Contents

Fig. 1 Cost optimization versus ROI maximization

Fig. 2 Cost optimization versus ROI maximization

Note that the net stock of an item equals its physical stock minus its backlog, and
the inventory position of an item equals the sum of its net stock and its outstanding
orders.
Contents 19

Table 2 Inventory system definitions, stochastic demand

Variable Definition
H holding cost rate per item on stock
p penalty cost rate per item short
ξ parameter of the inventory control policy
Pξ inventory control policy with parameter ξ
Xξ (t) net stock at time t under control policy Pξ
Yξ (t) inventory position at time t under control policy Pξ
Xξ long-run stationary net stock under control policy Pξ
C(ξ ) long-run average costs under control policy Pξ
C1 (ξ ) long-run sum of holding and penalty costs under control policy Pξ
C2 (ξ ) long-run average costs minus the long-run sum of holding and penalty costs
under policy Pξ
D(0,t] demand until time t
Qξ (0,t] cumulative amount replenishment orders received in stock until time t under
policy Pξ

3.2.1 Newsvendor fractile

In this subsubsection we show that under most control policies used for SISE models
the so-called Newsvendor equation holds under the optimal policy parameters. As
stated in section 1, the Newsvendor equation states that the non-stockout probability
at an arbitrary point in time equals the quotient of penalty cost rate and the sum
of holding cost rate and penalty cost rate. This quotient is called the Newsvendor
p
fractile and equals p+h . A theorem provides a set of sufficient conditions that the
cost structure and control policy structure must satisfy to yield this result. These
sufficient conditions are easy to verify.
The result holds for both continuous time and discrete time. In the continuous
time case we incur costs at the relevant rate per time unit per unit short or in stock,
while in the discrete time we incur costs at the relevant rate per item short or in stock
at the end of a time unit. The policy Pξ has a control parameter ξ and possibly other
parameters. We assume in the theorem below that the other parameters are constant.
For ease of reference we may think of Pξ as an (s, nQ)-policy, where ξ is the
reorder point s and the other parameter is the lot size Q. With the above notation we
can formulate the following theorem that provides a set of conditions under which
the optimal ξ , i.e. the value of ξ that minimizes the long-run average cost, satisfies
the Newsvendor equation .

Theorem 1. Assume that the inventory process Xξ (t),t ≥ 0, under Pξ is ergodic

∀ξ ∈ R. Assume that holding and penalty costs are the only costs affected by the
inventory process Xξ (t),t ≥ 0, and all other costs are affected by the replenishment
process Qξ (0,t],t ≥ 0. Assume that C(ξ ) is continuously differentiable in ξ . Let ξ ∗
denote the cost optimal value of ξ .
20 Contents

Assume that the following property holds ∀ξ , ε ∈ R,

Xξ +ε (t) = Xξ (t) + ε, ∀t ≥ 0 (Net stock translation property)

Then the following properties hold:

(i) Qξ +ε (0,t] = Qξ (0,t], ∀t ≥ 0
(ii) C2 (ξ ) = C2 (0), ∀ξ ≥ 0
(iii) C2 (ξ ) = HP{Xξ > 0} − pP{Xξ ≤ 0}, ∀ξ ∈ R
0

p
(iv) P{Xξ ∗ > 0} = p+H

Proof. By assumption we have that

Xξ +ε (0) = Xξ (0) + ε.

As the exogenous stochastic demand process D(0,t] is not affected by the inven-
tory control policy and the inventory balance equation states that net stock at time
t equals the net stock at time 0 plus cumulative orders received minus cumulative
demand, it follows that

Qξ (0,t] = Xξ (t) − Xξ (0) + D(0,t]

= (Xξ +ε (t) − ε) − (Xξ +ε (0) − ε) + D(0,t]
= Xξ +ε (t) − Xξ +ε (0) + D(0,t]
= Qξ +ε (0,t).

This proves (i). Then (ii) follows directly from the assumption that the costs
other than holding and penalty costs are affected only by the replenishment process.
This leaves us to proof (iii) and (iv).
Using the definition of C2 we obtain

C2 (ξ ∗ + ε) −C2 (ξ ∗ ) = (E[HXξ++ε + pXξ−+ε ]) − (E[HXξ− + pXξ− ]). (7)

Let us consider an arbitrary point in time t. We distinguish between the events

{Xξ (t) > 0}, {−ε < Xξ (t) ≤ 0}, and {Xξ (t) ≤ −ε}. This yields
Contents 21

E[HXξ++ε (t) + pXξ−+ε (t)] = E[(HXξ++ε (t) + pXξ−+ε (t))1{Xξ (t)>0} ]

+E[(HXξ++ε (t) + pXξ−+ε (t))1{−ε<Xξ (t)≤0} ]
+E[(HXξ++ε (t) + pXξ−+ε (t))1{Xξ (t)≤−ε} ]
= E[HXξ +ε (t)1{Xξ (t)>0} ]
+E[(HXξ++ε (t) + pXξ−+ε (t))1{−ε<Xξ (t)≤0} ]
+E[−pXξ−+ε (t))1{Xξ (t)≤−ε} ]

Now we use the sample path property in the theorem, which leads to the follow-
ing equation.

E[HXξ++ε (t) + pXξ−+ε (t)] = E[H(Xξ (t) + ε)1{Xξ (t)>0} ]

+E[(HXξ++ε (t) + pXξ−+ε (t))1{−ε<Xξ (t)≤0} ]
+E[−p(Xξ (t) + ε)1{Xξ (t)≤−ε} ] (8)

In a similar way we can write

E[HXξ+ (t) + pXξ− (t)] = E[HXξ (t)1{Xξ (t)>0} ]

+E[(HXξ+ (t) + pXξ− (t))1{−ε<Xξ (t)≤0} ]
+E[−pXξ (t))1{Xξ (t)≤−ε} ]. (9)

Combining equations (8) and (9), taking the limit t → ∞, and using the definition
of the indicator function, we obtain an expression for the righthand side of equation
(7).

(E[HXξ++ε + pXξ−+ε ])
−(E[HXξ− + pXξ− ]) = HεP{Xξ > 0} − pεP{Xξ ≤ −ε}
+E[(HXξ++ε + pXξ−+ε )1{−ε<Xξ ≤0} ]
−E[(HXξ+ + pXξ− )1{−ε<Xξ ≤0} ].
(10)

Rearranging terms of equation (10) and using equation (7), we obtain,

C2 (ξ +ε)−C2 (ξ )
ε − (HP{Xξ > 0} − pP{Xξ ≤ −ε}) =
E[(HXξ++ε +pXξ−+ε )1{−ε<X ≤0} ]−E[(HXξ+ +pXξ− )1{−ε<X ≤0} ]
ξ ξ
ε. (11)
22 Contents

Now note that under the condition that Xξ ≤ −ε both Xξ and Xξ +ε have an abso-
lute value less than ε. Using this observation then this and equation (11) yields the
following inequality,

C2 (ξ + ε) −C2 (ξ )
| − (HP{Xξ > 0} − pP{Xξ ≤ −ε})| ≤ 2(H + p)P{−ε < Xξ ≤ 0}.
ε
(12)

As we assume that C2 (ξ ) is continuously differentiable, it must hold that Xξ has a

continuous distribution. Thus, taking the limit ε → 0, the righthandside of equation
(12) goes to 0. Then we find the following expression for the derivative of C2 (ξ ),

C2 (ξ ) = HP{Xξ > 0} − pP{Xξ ≤ 0}, ∀ξ ∈ R,

0
(13)

which proves property (iii) of the theorem 1. In order to find the cost-optimal ξ ∗
we only need to minimize C2 (ξ ) with respect to ξ . As the KKT condition for ξ ∗
should hold, we have

p
HP{Xξ∗ > 0} − pP{Xξ∗ ≤ 0} = 0 ⇐⇒ P{Xξ∗ > 0} = p+H . (14)

This concludes our proof. t

u
The theorem provides the underlying principle for the consistent emergence of
the Newsvendor fractile when analyzing stochastic single-echelon inventory sys-
tems. For all commonly used inventory policies, such as MRP’s time-phased order
point (cf. section 4.4), (s, S), (s, nQ), (R, s, S), and (R, s, nQ), it is easy to verify that
if the reorder point is increased by some ε, while keeping the lot size parameter
constant, i.e Q and S − s, then the net stock also increases by ε. The aforementioned
policies observe the inventory position Y (t) and if Y (t) is less than the reorder level
s an order is placed at the supplier. For the case of an (s, nQ)-policy or an (R, s, nQ)-
policy, as many times a quantity of size Q is ordered as needed to increase the
inventory position above the reorder level s. For the case of an (s, S)-policy or an
(R, s, S)-policy, the order quantity raises the inventory position to order-up-to-level
S. Inventory control policies that only use information on outstanding orders, in-
ventory and demand for the single item under consideration, are called installation
stock policies.
An important consequence of the theorem is that it holds for any multi-item
multi-echelon system, where the end-items are controlled according to (one of the
earlier mentioned) installation stock policies for which the net stock translation
property holds. I.e. for any end-item the Newsvendor fractile holds under that situa-
tion. The reasoning for this is as follows. Assume at time 0 some outstanding orders
and assume that in system 1 the net stock equals Xξ ,k (0) and in system 2 the net
stock equals Xξ ,k (0) + ε for each end-item k. Assume that the control parameter in
Contents 23

system 1 equals ξk and in system 2 ξk + ε. For the above-mentioned installation

stock policies it is easy to show that the orders generated over time towards the
upstream multi-echelon system are identical. As this upstream system’s response is
completely determined by the orders received from all end-items, it follows that this
response, i.e. the replenishment process, is also identical for system 1 and system 2.
It is easy to see that property (ii) is a necessary and sufficient condition for the net
stock translation property to hold. Thus the optimal installation stock policy must
satisfy the Newsvendor fractile.
It is worthwhile to note here that the theorem may also hold for the so-called ech-
elon stock policies discussed in section 4, provided the end-items are controlled by
installation stock policies. The echelon stock policies are using the echelon inven-
tory position of an item as state variable. The echelon inventory position is defined
as the physical stock of an item plus its outstanding orders plus the echelon inven-
tory positions of its parent items, taking into account how many items are needed to
make one parent item. It follows that for end-items the echelon inventory position is
the same as the net stock, whereby end-item echelon stock policies are in fact instal-
lation stock policies. It suffices to check the set of conditions for the end-items only
to see if the Newsvendor equations hold for end-items under echelon stock policies.
In subsection 4.8 we show that the Newsvendor equation result can be extended to
Newsvendor equations related to each item in the MIME system.
We also note here that we did not make any assumption on the demand pro-
cess, other than assuming the demand process yields a continuous differentiable
cost function C2 (ξ ). This implies that the Newsvendor fractile also holds for non-
stationary demand, such as seasonal demand and autocorrelated demand, provided
that the control policy used yields the net stock translation property. In case we as-
sume that the policy is driven by a demand forecast, such as is the case for rolling
scheduling policies, with a stationary additive forecast error, then a time-phased or-
der point policy is appropriate and this policy ensures that the net stock translation
property holds.
Finally we note here that for the case of discrete demand it can be shown that the
net stock translation property is a sufficient condition for the Newsvendor fractile
inequality to hold.
The natural emergence of the Newsvendor fractile points towards the importance
of the non-stockout probability as a performance measure. A target non-stockout
probability relates one-to-one to a penalty cost value. This provides insight into the
implicitly assumed penalty costs per unit per unit time. Another advantage of the
non-stockout probability as a service measure is the ease of determining its actual
value over some time period. It suffices to keep track of the inventory level. The
most-prominent service level in inventory management literature is the fill rate. In
practice it is often used, too, but a major issue is that computation of the fill rate
requires knowledge of the actual demand over the time period of interest. In most
situations demand data are not known, only sales data. As sales is impacted by the
inventory availability, the use of sales data may give a too optimistic picture of the
situation.
24 Contents

3.2.2 Sensitivity analysis

In the introduction we discussed the difference between inventory model results and
results in practice under the influence of planners and schedulers. One of the typical
consequences is that we do not know which inventory model is applied. We may
infer the model used from the demand data, inventory data and replenishment order
data. This would be an interesting subject for further research. Another question
that arises in this context is, whether it matters what inventory model is used to
determine the control parameters to be used in practice. To answer this question we
use the algorithms from De Kok (1991a) to determine inventory control parameters
for the (R, S)-model, the (R, s, S)-model, and the (R, s, nQ)-model.
We set up a computational experiment as follows. Knowing that the main cost
drivers of inventory management are the frequency of ordering and the average in-
ventory level, we fix the average lot size and determine the average inventory needed
under an (R, s, nQ)-policy with R = 1 to meet a target fill rate level P2 . The average
lot size may be larger than Q, as at some instances multiple Q’s may be ordered as
a consequence of a large demand causing a high undershoot of the reorder level s.
For an exact formula for the average order size in an (R, s, nQ)-model we refer to
De Kok (1991b).
Contents 25

Table 3 Sensitivity analysis of inventory models

26 Contents

Then we determine the (R, s, S)-policy with R = 1 and the (R, S)-policy that have
the same average inventory and average lot size. The first step in the procedure is
to determine (S − s) for the (R, s, S)-policy and R for the (R, S)-policy such that the
average lot size equals the one for the (R, s, nQ)-policy applied to the same instance.
Then we can compute s and S for the (R, s, S)-policy and (R, S)-policy, respectively,
such that the average inventory equals the one for the (R, s, nQ)-policy applied to
the same instance. For these policies we determine the actual fill rate P2 .
In Figure 3 we present the results. As expected, the fill rates of the (R, s, S)-policy
are slightly better than that of the (R, s, nQ)-policy and in most cases identical. It is
more striking that the fill rates under an (R, S)-policy are close to the (R, s, nQ)-
policy fill rates and get lower as lot sizes get larger. The impact of the constant
replenishment lead time L and squared coefficient of variation c2 are negligible.
Thus it can be argued that the choice of the inventory control model is not crucial and
one can choose for the model that is mathematically more tractable. In particular the
virtually identical results for the (R, s, S)-model and the (R, s, Q)-model favors the
latter model as its inventory position distribution can exactly be determined (uniform
distribution), while the inventory position distribution of the (R, s, S)-model involves
the renewal function of the demand distribution, which is intractable in most cases.
We would like to mention here that the findings presented here are confirmed by
repeated experimentation in different real-life situations.
Another argument for being cautious to assume that we should always choose
the cost-optimal (R, s, S)-model is that inventory models are abstractions from re-
ality. Firstly, it may be that the production process requires fixed lot-sizes, which
is captured by the (R, s, nQ)-policy. Secondly, production of replenishment orders
involves equipment and human resources, which availability is planned to process
multiple replenishment orders. For resource planning, timing of production order
starts is key. For both the (R, s, nQ)-policy and the (R, s, S)-policy, timing of replen-
ishment orders is stochastic, while timing is fixed under the (R, S)-policy. Given
the robustness of the inventory model performance, in particular for items ordered
at high frequencies, the (R, S)-policy is particularly suited for taking resource con-
straints into account, albeit implicitly.

3.2.3 Empirical validity

As with every model applied in practice, it is important to ascertain its empirical va-
lidity. After 30 years of extensive development and testing of these models in prac-
tice through numerous MSc thesis projects, it is safe to say that SISE have proven
their empirical validity. We refer to the previous section and the introduction for
modelling considerations, and in particular to the distinction between intervention-
dependent and intervention independent performance measures. Only the latter are
suited for validation of inventory models and for setting inventory model param-
eters in business information systems. The same MSc projects revealed that most
ERP systems have mathematically incorrect algorithms for setting inventory model
parameters. One of the most striking errors found is the application of the algorithm
Contents 27

for determining P1 , the non-stockout probability at the end of a replenishment cy-

cle, to the situation with a P2 (fill rate) performance target. Our only explanation
for these incorrect formulas in ERP systems is the widespread (if not exclusive)
use of the P1 -measure in basic textbooks on production and inventory management,
combined with the extensive use of the normal demand assumption, which is rarely
valid, due to the intrinsic high volatility of item demand.
Another typical mathematical anomaly in ERP system inventory models is the
modification of the algorithm for items with high demand volatility. In that case the
modified algorithm ensures that the inventory is lower than according to the correct
mathematical analysis. Still, there is a relevant underlying tacit knowledge aspect
to this type of modification. In this case we should understand what is the cause of
high item demand volatility and how it manifests itself. High demand volatility can
easily be understood when simulating a gamma distribution with a high coefficient
of variation. One observes many small demands and now and then a demand which
is many times the average demand. We should be aware that in practice demand
is transient. We introduce an item, at which moment the demand process starts.
Over time, demand increases, possibly stabilizes around some level during some
period of time, and decreases to zero after that. This product life cycle can last
from a few months (e.g. a fashion T-shirt) to many years (e.g. a white T-shirt). The
high inventory associated with high volatility demand results from the recurring
peak demands over an infinite time length. In practice we may face many small
demands and only a few peak demands. And we may even face no peaks at all, as we
derive demand volatility from past demand data. The peak demands are oftentimes
due to special events, which may well be known far in advance, or in other cases
customers understand it may take longer than normal to satisfy. These are reasons
to ignore the peak demands, leaving us with many small demands. These constitute
a demand process that lends itself perfectly for being managed with the inventory
control models discuss above. The remaining volatility without the peak demands is
much lower, whereby a lower average inventory is sufficient to satisfy these smaller
demands at the target performance level. We refer to De Kok (1993) and Dekker
et al. (1998) for further details.
We conclude that SISE inventory systems are empirically valid provided proper
modelling considerations are taken into account (cf. section 2). The mathematical
analysis of these models is available in the form of exact and accurate approximate
algorithms, which can easily be implemented in large-scale inventory management
systems. For extensive discussions of the mathematical analysis we refer to Zipkin
(2000) and Axsäter (2015).

4 Uncapacitated multi-item multi-echelon models

In this section we discuss the modelling and analysis of MIME systems without re-
strictions on the resources that perform the transformation of child items into their
parent item. We only consider the material constraints that are constituted by the
28 Contents

available child item availability at the moment of the order release of the parent
item. In subsection 4.3 we present a generic MIME model under the assumption
of constant lead times, for which we provide a motivation in subsection 4.1. We
discuss feasibility conditions that control policies for MIME inventory system must
satisfy in subsection 4.4. In subsection 4.5 we introduce two concepts that emerge
when extending the control of SISE inventory systems to MIME inventory systems:
synchronization and allocation. Synchronization ensures that orders are not released
too early, allocation ensures that material availability constraints are respected. We
first discuss these two concepts in the context of convergent MIME systems in sub-
subsection 4.5.1 and extend the concepts for application to general MIME systems
in subsubsection 4.5.2. We show that the control of general MIME structures can be
derived from associated divergent MIME structures, which we denote as decision
node structures. The base stock policies for the divergent MIME systems applied to
general MIME systems are called Synchronized Base Stock (SBS) policies. While
for convergent MIME systems the concepts provide insight into the structure of op-
timal policies, for general systems this is no longer the case. We provide insight on
this matter by discussing some structural properties of Bills of Material in subsub-
section 4.5.3. In subsection 4.6 we use the case example presented below in sub-
section 4.2 to illustrate the derivation of decision node structures for general MIME
systems. As we have reduced the control of general MIME systems to the control
of divergent MIME systems, we discuss optimal policies for the latter systems in
subsection 4.7 and close-to-optimal policies that are mathematically tractable and
easy to implement in practice. The extension of the Newsvendor equation for SISE
systems to divergent MIME systems is discussed in subsubsection 4.8. In subsection
4.9 we discuss the performance of SBS policies in comparison with rolling schedul-
ing policies is commonly used in practice to decide on order releases in MIME sys-
tems. In subsection 4.10 we discuss the empirical validity of SBS policies as derived
from extensive empirical research. The empirical validity implies that, likewise for
SISE systems, average item inventories and average item order frequencies deter-
mine end-item customer service. Finally, subsection 4.11 is devoted to the strategic
positioning of inventory capital across the supply chain.

4.1 Material availability and stochastic lead times

Though we did not explicitly discuss the role of lead times in the analysis of SISE
systems, all performance characteristics of these systems depend on the distribution
of the replenishment lead time (cf. Axsäter (2015) and Zipkin (2000)). If we assume
the replenishment lead time is constant, this can be seen as modelling the situation
where the supplier of the item ordered always can deliver the order according to a
pre-specified nominal lead time. This may be due to ample inventory at the supplier
or the capability of the supplier to produce and ship the item according to the nom-
inal lead time. If we learn from replenishment lead time data that the actual lead
time is stochastic, then apparently the supplier is not always capable of adhering to
Contents 29

the nominal lead time. This incapability may be due to lack of available item in-
ventory to ship from stock, when this is necessary to meet the nominal lead time.
It could also be due to lack of material availability, whereby production is delayed
and subsequently replenishment order shipment is delayed, which eventually leads
to exceeding the nominal lead time. Next to lack of inventory availability, insuffi-
cient resource availability in transportation, warehousing or production can cause
delays that imply exceeding the nominal lead time. Modelling explicitly finite re-
source availability is discussed in Section 5. In this section we concentrate on the
explicit modelling of inventory availability over time at stock locations supplying
other stock locations. We note here that finite resource availability can be modelled
implicitly by assuming a nominal production lead time that enables the production
department to smooth resource requirements within this lead time so that both ca-
pacity constraints are met and due date requirements derived from the order arrivals
and the nominal production lead times are met with high probability. In that way the
nominal production lead time decouples the inventory management across multiple
stockpoints from shopfloor management at each production location. For an exten-
sive discussion on this way of hierarchically decomposing planning and control in
supply chains we refer to de Kok & Fransoo (2003). For the remainder of this sec-
tion we thus assume that each item order is delivered according to its nominal lead
time, provided that the child items of the item are sufficiently available in stock to
release the order to production upon arrival.

4.2 The example supply chain

Throughput this section we use an example to illustrate the problems emerging when
developing control policies for general MIME systems. We consider a 2-echelon
supply chain consisting of 4 end-items, numbered 1,2,3, and 4, and three compo-
nents, numbered 5,6, and 7. End-item 1 is assembled from one item 5 and one item
6 with a nominal lead time of 2 periods. End-item 2 is assembled from one item 5,
two items 6, and two items 7, with a nominal lead time of 4 periods. End-item 3 is
assembled from one item 6 and two items 7 with a nominal lead time of 2 periods.
End-item 4 is assembled from one item 6 and one item 7 with a nominal lead time of
4 periods. The component items 5, 6, and 7 have a nominal lead time of 2, 10, and 6
periods, respectively. This structural information about the supply chain is depicted
in figure 3. The average demand per period for end-items 1,2,3, and 4 equals 100.
The example supply chain represents a general structure as we see that each
component can be associated with an embedded divergent supply chain, and each
end-item can be associated with an embedded convergent supply chain. Note that
items 6 and 7 have both one and two as quantity per in relation to the end-items.
As we will see different quantities per of child items in their parent items creates
additional complexity on top of the inherent complexity of interactions between
child item and parent items as a consequence of limited child item availability.
30 Contents

Fig. 3 The example supply chain

4.3 Modelling multi-echelon inventory systems

The challenge of multi-echelon inventory system modelling is to capture the in-

teractions between item orders across the inventory system. These interactions are
quite complex due to the stochasticity of demand over time (uncertainty-driven com-
plexity) and the structural complexity of the Bill-Of-Material (BoM). The Bill-Of-
Material describes what child items are needed and in what quantity, to produce one
particular item. In this section we aim to provide fundamental insight into the re-
lationship between order release decisions over time in real-life supply chains, i.e.
general MIME inventory systems. The emphasis is on constraints imposed by or-
der release decisions in the past on order release decisions now and in the future.
We provide necessary conditions for feasible order release decisions, and sufficient
conditions that enable transparency of the order release decisions over time. These
sufficient conditions are induced by MIME inventory control policies, for which we
can find optimal policies under an assumption that may not always hold, but whose
Contents 31

performance is sufficiently accurate for practical purposes. Furthermore these poli-

cies allow for highly efficient calculation of operational order release decisions and
allow for efficient computation of close-to-optimal policies. Thus this section cen-
ters around a particular perspective on controlling general MIME inventory systems,
while other approaches have been proposed in literature. We refer to these other
approaches where appropriate, so that we can articulate the main differences with
the MIME inventory control concept discussed below. In general we can state that
the other approaches concentrate on calculating safety stock norms without dealing
with the operational details of order release decisions over time in MIME inventory
systems, or only consider the operational details of order release decisions without
dealing with the question how safety stock norms must be set.
Let us define the BOM and some other relevant structural aspects of multi-
echelon inventory systems.

Table 4 Defining a multi-echelon inventory system structure

Variable Definition
ai, j number of items i required to produce one item of its parent item j
Vi set of immediate successors of item i
Ui set of all predecessors of item i
Wi set of all successors of item i and item i itself, the echelon of i
E set of all end-items
Ei set of all end-items in Wi

The matrix (ai, j ) is the BoM matrix. In practice the BoM matrix is sparse, as the
natural built-up of a product from materials into subassemblies and final assembly
creates only parent-child relations between materials and subassemblies, and be-
tween subassemblies and final assembly. In ERP systems only the BoM relations
are stored, which ensures efficient storage-and-retrieval. We assume that only end
items face customer demand. If customers also demand non-end-items, such as in
the case of spare part demand, then it is easy to modify the network structure by
introducing dummy end-items for each upstream item with customer demand. The
set Vi is the set of parent items of item i. The set Ui contains all the items from which
item i is assembled (or in general, produced). The set Wi is the set of items in the
echelon of i.

Case example
For our case example we have E = {1, 2, 3, 4}, W1 = E1 = {1}, W2 = E2 = {2},
W3 = E3 = {3}, W4 = E4 = {4}, E5 = {1, 2}, E6 = {1, 2, 3, 4}, E7 = {2, 3, 4},
W5 = {1, 2, 5}, W6 = {1, 2, 3, 4, 6}, W7 = {2, 3, 4, 7}, U1 = {5, 6}, U2 = {5, 6, 7},
U3 = {6, 7}, U4 = {6, 7}.

In reality company supply chains consists of three or more echelons. In Business

to Consumer (B2C) supply chains we identify materials sourced from outside sup-
32 Contents

pliers, finished goods at the factory and finished goods in the distribution network.
In Business to Business (B2B) supply chains we identify parts and subassemblies
at 1st tier suppliers and finished goods at the Original Equipment Manufacturer
(OEM). Furthermore many items have both multiple children and multiple parents.
For such an item i we have |{a j,i > 0}| > 1 and |{ai, j > 0}| > 1. Clearly, for an
end-item k, |{ak, j > 0}| = 0. The distribution network is a network where we have
|{a j,i > 0}| = 1 for all items, and |{ai, j > 0}| >= 1 for all items but the end-items
k. As we are interested in finding optimal control policies, it is important to be
aware that this is not possible for divergent MIME systems, as follows from Ep-
pen & Schrage (1981), Diks & De Kok (1998), and Doğru et al. (2009), amongst
others. This implies that for any realistic supply chain structure there is no hope
that we can find optimal control structures. This is an important observation that
should be taken as a starting point for studying multi-echelon inventory systems un-
der stochastic demand. If it is impossible to find optimal policies, the infinite space
of heuristic policies opens up as a relevant space, similar to the situation of NP-hard
combinatorial optimization problems. We should be aware that many mathematical
properties only hold under optimality, such as the KKT conditions, and the Bellman
equations. Though the Bellman equations cannot be used due to the curses of di-
mensionality, the KKT conditions can still be applied for finding optimal policies
within a class of policies, by formulating the optimization problem as a function
of the policy parameters. A nice example of such an approach for general systems
can be found in Ettl et al. (2000). Before we discuss the optimization of given poli-
cies, we discuss some concepts that are foundational for controlling multi-echelon
systems under uncertainty in demand.

4.4 Feasibility of order release quantities

Likewise any production and distribution planning problem, we want to balance

supply and demand. In principle supply should be such that all demand can be sat-
isfied according to the requirements of the customers. As demand is uncertain, this
is not possible, as decisions on supply, i.e. order releases, precede the revelation of
demand into customer orders. This also implies that order release decisions must be
based on demand forecasts. As it matters, an analysis of SISE systems that explic-
itly incorporates the demand forecasting process into the inventory control policy
leads to mathematical intractability. Thus most papers assume stationary demand
processes, where it is implicitly assumed that we have an optimal demand forecast,
as we assume to know the mean. In reality this is not the case.
Despite the mathematical intractability of SISE (and MIME) systems under de-
mand forecasting, it is possible to formulate necessary conditions that an inventory
control policy for MIME systems should satisfy. We emphasize here that these nec-
essary conditions concern the inventory control model. We start from the premiss
that the model as an abstraction of reality should be internally consistent. An im-
portant concept regarding internal consistence is feasibility of a decision: a decision
Contents 33

should be such that all constraints formulated are satisfied. As the model is an ab-
straction, it can be the case that decisions regarded as infeasible according to the
model, are feasible in reality. However, such a solution is assumed to surface out of
the interaction between a decision support system that proposes a solution and the
decision maker who sees possibilities to improve the solution by relaxing constraints
that are violated according to the model. This is common practice in planning and
scheduling, although it is not always considered as explicitly as formulated here. In
order to formulate the necessary conditions for feasible order release decisions in
MIME systems given in de Kok & Fransoo (2003), we assume that we must decide
on order release decisions at time 0. We define the following variables.

Table 5 Defining a multi-echelon inventory system operational characteristics

Variable Definition
Li nominal lead time of item i
c
Li,k cumulative nominal lead time of item i with respect to end-item k
Xi (t) net stock of item i at time t
νi safety stock of item i
Yi (t) echelon inventory position of item i at time t, immediately after ordering
Qi (t) order quantity of item i that replenishes inventory of item i at time t
Pi (t) quantity of item i produced in period (t − 1,t]
Zi (t) echelon stock of item i at time t
Dk (s,t] independent demand for end-item k during the interval (s,t]
Di (s,t) sum of independent demand for end-items k ∈ Ei during the interval (s,t]
Gi (s,t] dependent demand for end-item i during the interval (s,t]
ri (t) quantity released of item i at time t
Bi,k (t, s) item i coverage of future demand of end-item k from time t until time t + s

The cumulative lead time Li,kc is the sum of all nominal lead times over all items

on the path from item i to end-item k. Furthermore, the dynamic variables defined
above can refer to the instance of the planning problem to be solved at time 0, or to
the actual realizations over time. As we assume that the model should be internally
consistent, we assume that the dynamic variables refer to the solution of the planning
problem at time 0, unless stated otherwise.
The necessary conditions for uncapacitated MIME systems concern material
availability.

N
Gi (t − 1,t] = ∑ ai j r j (t) ≤ Xi (t), ∀i (15)
j=1

Noting that Gi (t − 1,t] denotes the dependent demand at time t, inequality (15)
states that the amount of item i needed to release the orders of its parent items
cannot exceed item i’s availability. As the system is uncapacitated we assume that
the orders released at time t arrive according to their nominal lead time, i.e.
34 Contents

Qi (t + Li ] = ri (t), ∀i (16)

Inequality (15) is the Achilles heel of most inventory control policies proposed in
literature for general supply chains. If SISE control policies are assumed in a multi-
echelon context, then either one must compute explicitly the upstream delays due
to lack of child item availability (cf. Ettl et al. (2000) and Kiesmüller et al. (2004))
or ignore delays by setting high target service levels at upstream stockpoints (cf.
Graves & Willems (2000). The latter typically yields MIME policies that are far
from optimal (cf. subsection 4.11). The former approach assumes FCFS allocation,
as SISE inventory control policies do. FCFS allocation is not optimal in a MIME
setting, where there is mutual dependency between child item availability, especially
when there is a high commonality degree, i.e. many child items are used in the same
parent items. As mutual dependency is mathematically complicated, most papers on
MIME systems concern serial systems and divergent systems, where there is only
one child item for each parent item.
Inequality (15) is also the Achilles heel of MRP I logic. Though this logic is by
far most used in practice, it does not mean it is a correct MIME logic. Though MRP
I stands for Material Requirements Planning, a better description of the logic would
be Material Requirements Generation. The explosion process translates the Master
Production Schedule (MPS), the end-item order release (or make) plan over time,
into requirements for each upstream item, by using the BOM gozinto quantities
(ai j ), the nominal lead time {Li }, safety stocks, and lot sizing policies (cf. Orlicky
(1975)).
The MPS can be interpreted as the scheduled receipts for an end-item within its
nominal lead time, and planned receipts for the end-item k outside its nominal lead
time. The explosion process uses the nominal lead times L. to offset planned receipts
(replenishments of inventory) into planned orders, which constitute the dependent
demand for child items to be able to start production of the orders released over
time. The planned receipts are derived from the so-called time phased order point
logic and lot sizing rules. Ignoring the lot sizing constraints yields the following
ordering logic,

Gi (t − 1,t) = ∑ ai j r j (t), ∀i
j∈Vi
t−1
Qi (t) = Gi (0,t + 1) + νi − ∑ Qk (s), ∀i, ∀t ≥ Li (17)
s=1

The above equations indeed generate requirements aimed at each level to have
the planned net stock equal to νi from time t + Li onwards. But generating require-
ments is something completely different than planning! The logic generates (gross)
requirements and planned orders, assuming that all order releases can be realized
according to the generated timing. After generation of the gross requirements, the
inconsistencies, or infeasibilities, regarding item availability over time are signalled
as past due and schedule-in. As lot sizes have decreased structurally over the last
Contents 35

four decades, demand volatility causes regularly that the safety stock target cannot
be met taking into account gross requirements and scheduled receipts. In that case
past due of the first planned order is reported, which effectively is a signal that the
immediate planned order cannot be executed according to the ”plan”. In that case we
face the issue discussed above: ad hoc coordination of order release across different
links in the supply chains. Interestingly, past due immediate orders occur irrespec-
tive of the amount by which the planned on-hand is below the safety stock target. If
planned on-hand is well above zero at the time of arrival of the immediate planned
order, this order could be reduced in size to take into account child item availability.
In that case there is no need for ad hoc coordination across different links. However
this is a manual process without much support from the MRP system. Organiza-
tionally, material planners are often dedicated to a subset of suppliers delivering
different items. The mutual dependency of child items through their parents implies
that past due orders require a complex coordination process among multiple child
items, and multiple parent items, executed by multiple planners, without adequate
system support. The complex coordination problem is discussed in more detail in
the next subsection.
As a final remark, it is intriguing that the MRP I logic dates back to 1963, while
George Dantzig introduced Linear Programming (LP) in 1947. Both MRP I and
LP formulate the planning problem as a deterministic problem. As inequality (15)
is linear in the decision variables, it can be incorporated in the planning problem
formulation (cf. de Kok & Fransoo (2003)). To date MRP I has not been replaced
by LP, even though running a large scale LP is feasible. Our hypothesis is that using
LP as a substitute for MRP I would greatly improve material planning, and reduce
the need for ad hoc coordination across links in the supply chain.

4.5 Synchronization and allocation

In single-item single-echelon (SISE) inventory systems balancing supply and de-

mand implies covering average demand during the lead time plus safety stock with
orders in the pipeline and on-hand stock. In a sense this is not different in balancing
supply and demand in multi-item multi-echelon (MIME) inventory systems. The in-
ventory position in SISE inventory systems represents the coverage of demand over
the lead time plus the review period. The classical control policies mentioned in
Section 3 use the inventory position as the state variable that determines whether to
order or not, and if so, how much. The echelon inventory position in MIME sys-
tems plays a similar role as a state variable, but it is harder to identify what demand
is covered over what period. To see this, let us first define the echelon inventory
position and echelon stock for an item i at an ordering epoch t, immediately after
ordering,
36 Contents

Zk (t) = Xk (t), ∀k ∈ E
Lk
Yk (t) = Xk (t) + ∑ Qk (t, s), ∀k ∈ E
s=1
Zi (t) = Xi (t) + ∑ Y j (t), ∀i ∈/ E
j∈Vi
Li
Yi (t) = Zi (t) + ∑ Qi (t, s), ∀i ∈
/E
s=1
(18)

4.5.1 Convergent MIME systems

Now let us first consider a convergent system. In that case the supply chain produces
a single product. This allows to express item units in end-item units, so that w.l.o.g.
we can assume that ai, j ∈ 0, 1. Then the item i echelon inventory position Yi (t) must
cover the end-item demand over item i’s cumulative lead time plus review period
(which we assume equal to 1), implying that the demand for end-item k satisfied
directly from on-hand end-item inventory over time period (t,t + Li,k c + 1] is at most

Yi (t). Given the coverage Yi (t) at time t, it follows that at time t + s, after a demand
Dk (t,t + s] has occurred, the remaining coverage of demand over time period (t +
c + 1] equals Y (t) − D (t,t + s]. Thus we find
s,t + Li,k i k

c
Bi,k (t + s, Li,k + 1 − s) = Yi (t) − Dk (t,t + s], ∀i (19)

Now suppose that at time t + s some item j must be ordered to cover demand
for end-item k over its cumulative lead time L j,k plus review period. This implies
that s = Li,k − L j,k . As it does not make sense to create a coverage of demand for
end-item k exceeding the coverage from item i it follows that

Y j (t + s) ≤ Bi,k (t + s, Lcj,k + 1) = Yi (t) − Dk (t,t + s], ∀(i, j) ∈ {Lcj,k ≤ Li,k

c
} (20)

Thus the natural order created by the cumulative lead times of items in the supply
chain implies a natural set of constraints that the item coverages of end-item demand
should satisfy. Each item order implies a series of constraints, determined by the im-
mediate coverage created at the time of ordering and the realizations of demand over
time, for the ordering decisions for items with shorter cumulative lead times. This
formalizes the obvious fact that decisions taken in the past constrain decision now.
This set of constraints is the basis for the results in Rosling (1989), who proves that
echelon base stock policies are optimal for convergent systems and each convergent
system has an equivalent serial system. The echelon base stock policies for serial
system operate as follows: each point in time t the inventory position of item j is
increased to the minimum of its base stock level S j and the sum of its current eche-
Contents 37

lon inventory position plus the on-hand inventory at its predecessor. For convergent
systems the echelon base stock policies operate as follows: each point in time t the
inventory position of item j is increased to the minimum of its base stock level S j
and the minimum of the coverages Bi,k (t,t + L j,k + 1]. Because the coverages over
time constrain future coverages, the minimum of the coverages equals the coverage
of the item with the next longer cumulative lead time. So once the base stock levels
are known, the execution of the echelon base stock policies is highly efficient, if not
trivial.
The above may be somewhat technical. Conceptually, inequalities (20) constitute
the synchronization of orders in convergent supply chains. Synchronization can be
interpreted as the horizontal coordination of orders for components and subassem-
blies that together enable the assembly of the final product to be sold to the market.
Synchronization takes into account constraints set by earlier ordering decisions, im-
plying that it may be suboptimal to execute the item’s ordering policy based only on
its own state, i.e. its Y(t). The consequence of ignoring the constraints expressed in
inequalities (20) are stocks on hand that serve no purpose. We call these stocks dead
stocks, as it can be shown that these stocks do not contribute to the service of the
supply chain to the market. Below we discuss the calculation of these dead stocks.
At this stage it is of interest to return to the MRP I concept, i.e. Materials Re-
quirements Planning (cf. Orlicky (1975)). MRP I is a logic that generates orders in
a MIME inventory systems. The synchronization equations (20) are not respected
by the explosion equations (17). This implies that items are ordered too early and
are held in stock while waiting several time units for other items to arrive. The in-
formation to prevent this is available in the inventory database of the ERP system,
but not used.
Not respecting the synchronization constraints (20) and thereby ordering to early
is one thing, but ordering a quantity that is impossible to fulfill is another. We must
be aware that for each item we can distinguish between synchronization equations
that express that an order should not exceed coverage constraints to prevent order-
ing too early, and synchronization equations that express that an order cannot be
released by lack of upstream materials. The latter subset is given by

Y j (t + s) ≤ Bi,k (t + s, Lcj,k + 1), ∀i ∈ {m|am, j > 0}. (21)

The inequalities (21) for convergent MIME systems are equivalent to inequalities
(15), the feasibility constraints for order release quantities for any BoM structure.
Inequalities (21) represent the constraints set by the child items of an item j. They
are an example of so-called allocation constraints. Note that inequalities (21) repre-
sent explicit relations between order release decisions for different items over time,
while inequalities (15) represent implicit relations between order release decisions
over time. Such implicit relationships are typical for mathematical programming
formulations. While this can be seen as the strength of mathematical programming:
decomposing a problem in sets of constraints, for analysis of systems under uncer-
tainty we need explicit relationships in order to develop probabilistic expressions
38 Contents

for the performance of MIME systems, similar to those for SISE systems under
uncertainty.

4.5.2 General MIME systems

So far we restricted ourselves to convergent systems, for which the synchroniza-

tion inequalities (20) are unambiguously defined. For general BoM structures this
is no longer the case. To see this, suppose that item i has multiple parent items and
these parent items have multiple child items. Then there are many alternative ways
to allocate child item availability among these parent items. Assuming that some
allocation decisions have been taken before deciding on the allocation of item i
availability, we can formulate a set of synchronization equations related to the child
item allocation decisions already taken, but not for those child items which avail-
ability is allocated after the allocation decision for item i is taken. So the order in
which allocation decisions are taken determines the synchronization inequalities to
be respected, and the allocation decisions have many feasible solutions, likewise
inequalities (20) allow for many feasible solutions. In a way we should take these
decisions simultaneously, such that e.g. long-run costs average costs are minimized.
Without going into detail here, we note that formulation of the Bellman equations
for the optimal control problem related to MIME inventory system control shows
that we are faced with the curses of dimensionality, implying there is no hope for
ever finding an optimal solution to this problem. The interaction between alloca-
tion and synchronization decisions to be taken simultaneously could be seen as a
chicken-and-egg situation or, from a somewhat more pessimistic perspective, as a
catch-22 situation.
In De Kok & Visschers (1999) a possible way out of this situation is proposed,
which we conceptually discuss below in detail as it provides a deeper insight in
the fundamental complexities of inventory management in general MIME inventory
systems. The cutting of the Gordian knot of the general MIME control problem can
be formulated as: allocation before synchronization. The control policy proposed,
denoted as Synchronized Base Stock (SBS) policies in de Kok & Fransoo (2003), is
based on two structural characteristics of the MIME system: the BoM that defines
how item coverages relate to each other through the end-items that contain them,
and the item cumulative lead times that determines which order release decision
precedes the other. This structural foundation translates into a mapping from the
original BoM and associated lead times to a set of divergent decision node struc-
tures. These divergent decision node structures can be considered as the general-
ization of the serial system structures that are equivalent to convergent systems as
shown by Rosling (1989). Each decision node (C, EC ) represents a unique combina-
tion of a set of items C and a set of end-items EC that use these items and a point in
time that one or more items not in C are ordered. At this moment it is decided if the
coverage constraints of the items in C will be binding for the orders of the items not
in C, or not. If the latter is the case, an immediate overage of availability of items in
C is created that translates in physical inventory not used at the moment of receipt
Contents 39

of these items in C in their stockpoints. This immediate overage is represented as an

inventory level in the decision node associated with C and EC .
The leaves of a divergent decision node structure are 1-1 related to end-items,
which together constitute a subset of E. Each end-item is a leaf in exactly one di-
vergent decision node structure, but there can be multiple divergent decision node
structures. The root node of a divergent decision node structure relates to one or
more items that are released at some time t and no preceding order release decisions
have been made that relate to the end-items that contain the items released at time
t. Thus, the root node of the decision node structure contains the longest lead time
items of all end-items in the leaves of the divergent structure. More precisely, if
items i1 and i2 are in the root node of a decision node structure defined by the item
set C0 and end-item set EC0 , then we have

Lic1 ,k = Lic2 ,k ∀k ∈ EC0 (22)

This implies that all items associated with the root node of a decision node struc-
ture have the same cumulative lead time Li kc for all end-items associated with the
structure. Divergence in a decision node structure occurs when at some time t + s
an item is ordered that is contained in a strict subset of the end-items for which
items have been ordered before time t + s. The divergent decision node structure
unambiguously determines end-item demand coverages by items based on their cu-
mulative lead times. If at time t + s an item is ordered that causes divergence of the
decision node structure it belongs to, then the coverage of demand for end-items
during the cumulative lead time from time t + s until the moment of replenishment
of the end-items is first allocated between the subset of end-items that emerge due to
the divergence, and after that the relevant demand coverage constraints can be cal-
culated, similar to above for convergent structures. We refer to De Kok & Visschers
(1999) and de Kok & Fransoo (2003) for further details.
Under the modelling assumptions discussed above we can derive the order re-
lease decisions for each item in the MIME system at each point in time from the
ordering decisions in the decision node structures. If an item is represented in mul-
tiple divergent decision node structures, we simply sum the order release quantities
from each divergent structure. The SBS policies assume that echelon base stock
policies are used for each decision node, and an allocation policy is defined that
determines the amount released to each successor in case an item has insufficient
availability to satisfy the parent items’ orders. The latter case occurs if the echelon
stock at (C, EC ) is smaller than the sum of the echelon base stock levels of its suc-
cessors. If all orders can be satisfied, than the physical stock remaining at (C, EC )
represents the immediate overage created in the pipeline or on-hand of the items in
C in the original MIME system.
The decision node structures provide fundamental insight in the mutual depen-
dency of order release decisions taken over time in a MIME inventory system. Each
divergent tree represents a strict hierarchy in decision making. Whereas in the origi-
nal MIME inventory systems we are confronted with the above-mentioned catch-22
40 Contents

of mutual dependencies, under the SBS policies based on the principle of allocation
before synchronization, we have disentangled the problem mess, caused by mutual
dependencies of decisions in general MIME systems. This disentangling process is
demonstrated in section 4.6.

4.5.3 Commonality of items

In De Kok & Visschers (1999) it is assumed that the BoM matrix (ai, j ) satisfies
ai, j = 0, 1, ∀i, j. If this is not the case then the above described procedure can still
be applied, but some preprocessing is required. As stated above we represent the
assembly structure as a set of divergent decision node structures, and each node is
represented by item set C and end-item set EC . Let us assume that |EC | > 1 and
define bi,k as

bi,k := number of items i required to produce one item of end-item k, k ∈ E

The coefficients bi,k are often referred to as the Flat BoM coefficients.
As the divergent structure also presents the way the end-item demand propagates
upstream and a decision node is represented only by the item set C, demand propa-
gation must be the same for all items in C, up to a multiplicative factor, i.e.

bi1 ,k
= ci1 ,i2 , ∀k ∈ EC , ∀i1 , i2 ∈ C (23)
bi2 ,k

Clearly, this condition is not always satisfied in practice. In that case we cannot
combine items i1 and i2 in a single decision node. Assuming that i1 has the longest
lead time, this implies that upon ordering i2 , the decision node structure diverges
from the decision node containing i1 , defined by the sets C and EC , into multiple
decisions nodes containing i1 and i2 that satisfy equation (23). So even if i2 is used
by exactly the same end-items as i1 , we do not exploit the commonality in demand
for i1 and i2 , using the portfolio effect from adding demands for all end-items in
EC . Let us argue that we cannot exploit commonality like we might expect in case
equation (23) does not hold.
Towards this end let us first assume that equation (23) holds. In that case it is
easy to see that the propagated demand for i1 and i2 over time interval (s,t] satisfy

Di1 (s,t] = ci1 ,i2 Di2 (s,t] (24)

This implies that, up to a multiplicative constant, the demand processes for i1

and i2 are identical. Thus we can use the same control policies for both items over
the uncertainty period they have in common, defined by their cumulative lead times.
However, if equation (24) does not hold for i1 and i2 then the demand processes are
clearly not identical and thereby we cannot use the same inventory control policy.
Contents 41

The divergence of the decision node structure at the ordering of i2 is a structural

expression of this fact.

Table 6 Impact of heterogeneity in flat BoM on upstream demand heterogeneity

In Table 6 we illustrate the above by a numerical example. We consider two items

5 and 6 that are common to end-items 1-4. Item 5 has b5,k = a5,k = 1 while item 6
has b6,k = a6,k = k. We see that the demand for item 5 exhibits the portfolio effect: its
coefficient of variation cD is lower than that of the end-items. However, the demand
for item 6 has a coefficient of variation exceeding that of the fast mover, end-item 1.
This difference in demand behaviour must be reflected in different control policies
for items 5 and 6. While a demand of 50 for item 1 and 30 for item 2 results in the
same demand for item 5 as in the situation with a demand of 30 for item 1 and 50
for item 2, this is not true for item 6. Item 6 is sensitive to demand mix changes.
Even though the control policy resulting from the control policy for the diver-
gent decision node structures is not optimal (commonality cannot be exploited if
equation (23) is not satisfied, items are allocated to covering future demand for spe-
cific end-items too early), it does reflect the complexity induced by BoM structures
not satisfying equation (23). The author was first confronted with this phenomenon
when studying the commonality structure of TV’s at Philips Consumer Electronics
in 1990, finding that 100% common items exhibited high coefficients of variation in
demand, where at first sight low ones were expected. Looking at the bi,k coefficients
for these items revealed that these were different for different end-items k. Our men-
tal point of reference is typically bi,k = 1, from which we mentally add up demands
for end-items, which yields the expected commonality. However, heterogeneity of
the flat BoM coefficients bi,k drives demand volatility of upstream items.
Another common feature of BoM structures is that the same item, i say, is used
multiple times at different levels in the BoM. This implies that for item i at different
points in time an order release decision must be taken, creating coverage for future
demand of end-items over different cumulative lead times. In the MRP logic this is
dealt with by low level coding, where gross requirements for the same period are
consolidated before an order release decision is taken. By consolidating these deci-
sions, the fundamental relationship between the timing of orders for item i and what
part of an order covers which cumulative lead time demand of which end-items is
42 Contents

blurred. In the SBS policy approach this relationship is exposed by considering item
i at different levels of the BoM as different items. In that way we create a new BoM
structure, which is again translated into an associated decision node structure, from
which we derive feasible order release decisions, and from which we can directly
identify the relationship between order release decisions for different items over
time.
When assuming periodic review echelon order-up-to-policies for the divergent
decision node structures, we can consolidate the ordering decisions for each item
into base stock policies for the original general multi-item multi-echelon inventory
system. These base stock policies satisfy the material availability constraints (15).
However, in case of fixed lot sizes, this is no longer obvious. Though we can take
into account lot sizing constraints for the ordering policies in the divergent deci-
sion node structures, adding the resulting order releases may violate the lot sizing
constraints in the original system. The decision node structures allow for different
review periods for different items, yet at this moment in time it is even unclear how
to analyze divergent systems with different review periods. Though nested policies
have been studied for serial systems (cf. van Houtum et al. (2007)), it is argued
in Karaarslan et al. (2013) that nested policies for the decision node structures are
inappropriate in most cases: there is a strong positive correlation between the cost
of an item and its lead time. As high variable costs imply small lot sizes and thus
frequent ordering, optimal policies for controlling general multi-item multi-echelon
systems are unlikely to be nested.
Having said this, some of the issues that come with non-nestedness and lot-sizing
can be overcome by smart modelling. Firstly, large lot sizes are typical for relatively
cheap items. A large lot size of an item decouples the supply chain producing the
item from the supply chain using the item. This may imply that we can assume that
the item is always available without the need for excessive inventory. Secondly, we
can synchronize end-item coverages of short-lead-time items with large lot sizes
with the known end-item coverages of long-lead-time expensive items, thereby en-
suring that the short-lead-time-items are always available. This idea has been ex-
plored in Karaarslan et al. (2013) for a two-component-one-end-item supply chain.
Clearly more research is needed to test these hypotheses. In the remainder of this
section on uncapacitated MIME systems we assume that all items are controlled by
echelon-order-up-to policies with the same review period. For an extensive discus-
sion on modelling MIME systems we refer to Willems (2008). In section 4.10 we
discuss the modelling of MIME systems to test empirical validity of the quantitative
results derived from applying SBS policies.

4.6 Decision node structure for the case example

We use our case example to explain the creation of the decision node structure asso-
ciated with an assembly network. In the first step we create the decision node struc-
ture associated with each end-item. This yields serial structures in accordance with
Contents 43

the results in Rosling (1989). We use our decision node structure notation where
above each decision node (triangle) we show the associated end-item set and below
each decision node we show:
• its component set,
• the multiplicity of each item in the component set
• its added cost.

Fig. 4a Item 1: BOM and equivalent serial structure

Fig. 4b Item 2: BOM and equivalent serial structure

Fig. 4c Item 3: BOM and equivalent serial structure

44 Contents

Fig. 4d Item 4: BOM and equivalent serial structure

In the next step we consider the serial structures of items 1 and 2. In both struc-
tures item 6 is ordered first, at time 0, say. At time 4 item 7 is ordered, which is
used in end-item 2, but not in end-item 1. If we coordinate the control of the supply
chains of end-items 1 and 2, this implies that after 4 periods we must allocate the
coverage of future demand between item 1 and item 2, such that the coverage of
item 7 can be synchronized with the allocated coverage of item 6. At time 8 item
5 is ordered, upon which item 5’s coverage of end-item 1 demand is synchronized
with the coverage of item 6, and in parallel item 5 coverage of end-item 2 demand
is created by synchronization with the joint coverage of item 6 and 7 (as a result of
synchronization of these items’ coverages created at time 4). At time 10 both item
1 and 2 are ordered, implying that the synchronized coverages of their child items
can be used to release orders satisfying the material constraints defined by equations
(21).
Another issue must be dealt with before we arrive at the combined structure.
Root item 6 is used once in item 1, but twice in item 2. In order to create a consistent
combined decision node structure, we normalize the usage of root item 6 to 1 in both
end-items. This can be realized by redefining the unit of demand of item 2, which
relates to the observation leading to equation (24). The new unit of demand for item
2 is half of the original unit of demand for item 2. This implies that we must multiply
the end-item 2 demand per time unit by two. As item 7 is used twice in the original
item 2, it is used once in the modified item 2. Concerning item 5, it is used once in
item 1, but it is used only 0.5 times in terms of the new demand unit of item 2. Also
the multiplicity of the end-item itself is now 0.5, as one root item 6 is consumed by
half of the new demand unit of end-item 2. Thus we find the multiplicities for each
decision node in figure 5. From the multiplicities we can derive the cost added in
each decision node, E.g. when ordering item 5 and synchronizing its coverage with
that of items 6 and 7, 0.5 units of item 5 are added to the units of 6 and 7, which
implies an added cost of 0.5. Similarly, the added cost associated with the decision
node defined by item 2 equals 0.5, too.
We note here that the modification of the demand unit for item 2 is not depicted
in 5. The original demand unit can be derived from b6,2 = 2, i.e. the multiplicity of
the root item 6 in end-item 2.
Contents 45

Fig. 5 Decision node structure item 1 and 2

The next step is to combine the serial decision node structure of end-item 3 with
that of end-items 1 and 2. In the latter structure we see divergence at time 4 due
to the ordering of item 7 that is not used in end-item 5. Item 7 is used in item 3
together with item 6, however not in the same ratio as item 6 and 7 are in item 2.
Item 7 and item 6 are used twice in the original demand unit of item 2, but they
are used once and twice, respectively, in item 3. Thus item 3 cannot be added to
the existing decision node defined by end-item set {1, 2} and component set {6, 7}.
A new decision node is created, defined by end-item 3 and component set {6, 7},
where the multiplicity of 7 equals 2. At time 10 item 3 is ordered and uses the
synchronized coverage of items 6 and 7. As the multiplicity of 7 equals 2, the added
cost of the decision node associated with end-item set {3} and component set {6, 7}
equals 2. The added cost of the decision node associated with end-item set {3} and
component set {3} equals 1.
46 Contents

Fig. 6 Decision node structure item 1, 2, and 3

In the final step we combine the serial decision node structure of item 4 with that
of items 1, 2, and 3. Item 4 uses items 6 and 7 in a 1-1 ratio. This implies that item
4 can be added to the decision node defined by end-item set {2} and component
set {6, 7}. As item 4 does not use item 5, at time 8 we allocate the synchronized
coverage of items 6 and 7 for items 2 and 4 among these two items, so that item
5 can be synchronized with the allocated coverage for item 2. Thus a new decision
node is created that is defined by the end-item set {4} and the component set {6, 7}.
As the predecessor of this decision node has the same component set, the added
value of this node equals 0. At time 10 item 4 is ordered such that it respects the
synchronized coverage constraints set by items 6 and 7. Figure 7 depicts the decision
node structure for our case example. We have obtained a single divergent decision
node structure, as the root item of the structure is in all end-items.
Contents 47

Fig. 7 Decision node structure item 1, 2, 3, and 4

4.7 Control policies for divergent MIME systems

As SBS policies yield divergent decision node structures, finding an optimal SBS
policy for a general MIME systems reduces to finding an optimal policy for diver-
gent MIME systems. In Diks & De Kok (1998) the results of Clark & Scarf (1960)
for serial systems with linear holding and penalty costs have been generalized to
divergent MIME systems:
• Echelon base stock policies are optimal
• Optimal echelon base stock levels can be recursively determined, solving single-
variable equations
The results in Diks & De Kok (1998) only hold under the so-called balance as-
sumption defined by Eppen & Schrage (1981), which in turn is equivalent (for our
optimization problem) with allowing for negative allocations. If we do not allow for
48 Contents

negative allocations, then finding the optimal policy for divergent MIME systems is
as complex and computationally inhibitive as solving the Bellman equations for any
stochastic dynamic programming problem (cf. Doğru et al. (2009)). The cost per-
formance of a divergent MIME system computed under the balance assumption is a
lower bound of the cost performance under the true optimal policy. Several numer-
ical studies on divergent MIME systems (e.g. Federgruen & Zipkin (1984), Diks &
De Kok (1999), and Doğru et al. (2009)) have shown that the policies found under
the balance assumption, corrected for negative allocations in a discrete event sim-
ulation setting, have a cost performance close to this analytically computed lower
bound performance when the following conditions hold:
• End-item coefficient of variation of demand is below 1
• End-item mean demands are similar
The above conditions have deliberately been formulated informally, as the vari-
ous numerical studies show that it is impossible to derive a precise range in which
the balance assumption is effectively non-binding.
In their proof of the optimality of echelon base stock policies for divergent sys-
tems, Diks & De Kok (1998) show that the results only hold under optimal allocation
policies. However, these optimal allocation policies are intractable. This motivated
Van der Heijden et al. (1997) to study so-called linear allocation rules. These are
defined by allocation fractions qi j with j ∈ Vi which sum up to 1. In case the cu-
mulative successor item order release quantity exceeds item i available stock, the
excess is allocated according qi j and the allocated excess is subtracted from the
original successor item order release quantity. This leads to the following simple
control rule for divergent systems.

Y j (t) = S j − qi j ( ∑ Sm − Zi (t))+ , ∀ j ∈ Vi (25)

m∈Vi

The linear allocation policy ensures feasibility of the order releases, but suffers
from the same problem as the ”optimal” allocation policy under the balance as-
sumption: it may result into negative order release quantities. To address this issue,
Van der Heijden (1997) proposed to determine allocation fractions that minimize
(a proxy of) the probability that a negative allocation occurs. In de Kok & Fran-
soo (2003) the closed-form expression proposed by Van der Heijden (1997) was
corrected for a minor error, yielding the following expression for qi j .

σ (D j )2 E[D j ]2
qi j = 2
+ (26)
2 ∑m∈Vi σ (Dm ) 2 ∑m∈Vi E[Dm ]2

Thus ensuring a low probability of negative allocations requires to allocate a

larger part of the excess to successor items with high average demand and high de-
mand volatility. Though Van der Heijden et al. (1997) show that the so-called Bal-
anced Stock rationing improves the quality of the approximations of MIME system
Contents 49

performance measures, such as average inventories and fill rates, it does not remedy
the deterioration of performance in high demand volatility scenarios. Here it seems
that the only remedy is to keep more stock of items to reduce the occurrence of
excesses of demand over item availability. The computational study of Doğru et al.
(2009) shows that for high demand volatility scenarios the true optimal policy keeps
more inventory upstream than the policy derived under the balance assumption. It
is to date unclear how to compute echelon base-stock levels that circumvent the
inherent mathematical flaw implied by the balance assumption.

4.8 Generalized Newsvendor equations for divergent MIME

systems

Efficient computation of optimal control policies for divergent systems under the
balance assumption is based on the results from Diks & De Kok (1998). They show
that under the optimal policy so-called generalized Newsvendor equations hold. In
order to formulate these equations, we explicitly define the cost structure of a MIME
inventory system, which is in line with the cost structure for the SISE inventory
system in Section 3.

Table 7 Cost structure of a MIME system

Variable Definition
Hi holding cost rate per item i on stock
hi echelon holding cost rate per item i on stock
pk penalty cost rate per end-item k short
Iki (t) net stock of end-item k in the subsystem defined by Wi
Si echelon base stock level of item i

Then Diks & De Kok (1998) show that the optimal policy satisfies the following
optimality equations.

pk +Hi −hi
P{Iki ≥ 0} = pk +Hk , ∀k ∈ Ei . (27)

Recall that Ik i denotes the net stock at end-item k in the MIME system Wi of
which item i is the root node. From the divergent structure of the subsystems it fol-
lows that equations (27) can be solved recursively. Suppose we number the nodes
(stockpoints) in the divergent system by low level coding, similar to low level cod-
ing in MRP. I.e. the end-nodes of the divergent system have low level code 1, the
predecessors low level code 2, etc., and if a node has successors with different low
level codes, we number this node as the maximum low level code of the successors
plus 1. We start with solving equations (27) for the end-nodes with low-level code
50 Contents

1. These are in fact the same equations as for SISE systems, and provide the optimal
(echelon) base stock levels for each end-node (end-item). Then we solve equations
(27) for the nodes with low level code 2, etc. Finally we compute the base stock
level for the root node of the divergent system, which can be written as

pk
P{Xk ≥ 0} = pk +Hk , ∀k ∈ E. (28)

In Diks & De Kok (1998) is it proven that under the optimal policies the solution
to the |Ei | equations (27) is a single echelon base stock level for each item (node)
i. And equations (28) show that under optimal base stock levels and optimal alloca-
tion functions the Newsvendor fractile holds for the end-items. Remember that this
holds for MIME systems where the end-items are controlled by an installation stock
policy, too, through Theorem 1.
As the optimal allocation policies are intractable, Diks & De Kok (1999) propose
to use the linear allocation rules defined above. Although these policies do not sat-
isfy equations (25), as we have |Ei | equations for a single variable Si , the heuristic
based on averaging the |Ei | solutions for the echelon base-stock level performs well.
The recursive solution of equations (27) under the assumption of linear allocation
rules as described in equations (25) can be solved by bisection, which ensures that
the computation time is linear in the number of nodes in the decision node struc-
ture. As the maximum number of decision nodes arises when all end-item decision
node structures cannot be combined, the maximum number of bisection equations
is bounded by the product of the number of end items |E| and the total number of
items. The expressions for the non-stockout probabilities in optimality equations
(27) are mathematically intractable, but can be accurately approximated by two-
moment fits using mixed-Erlang distributions (cf. De Kok (2003)). The two-moment
fits are recursively applied to the shortfalls Zi (t) that emerge from the linear alloca-
tion policy equations (25). Thus, it follows that the computational effort involved in
finding close-to-optimal SBS policies is of the same order of magnitude as solving a
SISE model for each item in the MIME system. The numerical procedure sketched
here is used to determine the parameters of the SBS policies of which we present
the performance in the next section.
Before closing this subsection on divergent MIME systems, it should be pointed
out that there is an extensive literature on divergent MIME systems, which devel-
oped between 1980 and 2000. We refer to Axsäter (2003) for an excellent overview
of this literature. We chose to limit our discussion to divergent MIME systems under
periodic review, as our aim is to describe the operational control problems that come
with MIME inventory systems. Most other approaches are limited to two-echelon
systems and focus on the computation of control policies, and do not pay much at-
tention to the allocation problem by assuming e.g. FCFS of replenishment orders
from parent items.
Contents 51

4.9 Performance of SBS policies

The SBS policies have a structural foundation in the divergent decision node struc-
tures, and a control policy foundation in the base stock policies that are optimal
for serial MIME systems according to Clark & Scarf (1960), for convergent MIME
systems according to Rosling (1989) and for divergent MIME systems under the
so-called balance assumption according to Eppen & Schrage (1981), Federgruen &
Zipkin (1984), and Diks & De Kok (1998). All these results have been derived under
linear holding and penalty costs. In de Kok & Fransoo (2003) an example is given
that shows that SBS policies are non-optimal in general, yet their numerical com-
parison against LP-based rolling schedule policies shows that SBS policies yield
substantially lower long-run average costs than a policy that in practice is believed
to be effective for real-life large scale systems (i.e. LP under rolling scheduling).
The main findings in de Kok & Fransoo (2003) were as follows:
• Even though safety stocks for components were set to zero, LP-based control
yielded too high average inventories upstream. This is due to the fact that up-
stream items are cheaper than end-items, so that LP solutions favor to hold stock
in upstream items after periods of (coincidentally) low demand.
• End-items with identical BOM structures (including item lead times), identi-
cal cost built-up over time and identical demand processes could have widely
different safety stocks. This is mainly due to tie-breaking rules as under such
a situation the LP problem has infinitely many optimal solutions. This result
implies that under identical safety stocks similar end-items would have widely
different service levels. This phenomenon has been observed in practice in the
context of order release plans proposed by LP-based Advanced Planning and
Scheduling (APS) system engines. This could be a reason why APS system
planning proposals are often overwritten by planners.
• Similar to the last finding, in case an end-item has slightly lower holding costs
than others, it needs much higher safety stocks than the other items. This is due
to the extremal solutions generated by LP. In case of overages pushed down-
stream, the largest part (if not all) of the overage is allocated to the cheapest
end-item, and in case of underages the cheapest end-item is likely to get no
child item availability allocated at all. This yields a lumpy supply process for
the cheapest end-items, which effectively implies that end-item order releases
are delayed substantially or not. To compensate for the high end-item lead time
volatility, high safety stocks are needed to ensure the required service level.
From the above we concluded that LP tends to use fixed priorities in its alloca-
tion mechanism, which were shown to be ineffective in De Kok & Visschers (1999).
In their simulation study they explored the impact of the allocation before synchro-
nization. Assuming that it is always better to allocate as late as possible, they applied
simple allocation rules based on fixed priority, random priority and run-out times.
Under these rules, first the priority of end-items is determined and second the avail-
able child items are allocated end-item by end-item according to this priority. They
found that fixed priorities were performing worst, and the run-out time rule per-
52 Contents

formed best. But the main finding was that the postponement of allocation to the
very last moment, i.e. at the moment of order release, using the run-out time rule,
hardly improved the customer service levels of the end-items. This finding sug-
gested that the synchronization of item order releases over time is key for supply
chain performance. The results in de Kok & Fransoo (2003) confirmed this find-
ing as LP-based control allows for postponement of allocation to the last possible
moment.
In Spitter (2005) the comparison of SBS policies against standard rolling schedul-
ing was further explored. With the findings above in mind the rolling schedule
heuristics proposed aimed at preventing the extremal allocations that make LP per-
form badly under demand uncertainty. Two approaches were identified as more ef-
fective than LP-based rolling scheduling:
• Adding the linear allocation rules defined by equations (24) as constraints
• Replacing the linear objective function by a quadratic objective function
In the sequel we refer to the rolling schedule policy for which at the start of each
period we solve the standard LP model taking into account the feasibility constraints
(15) and (16) as the LPst -policy, the rolling schedule policy for which we add the
linear allocation constraints (24) to the standard LP model as LPalloc -policy, and the
rolling schedule policy based on the model with the feasibility constraints (15) and
(16) and a quadratic objective function as QP-policy. The linear objective function
takes into account the sum of holding and penalty costs over the planning horizon.
When exploring the appropriate formulation of the quadratic objective function, we
found that it sufficed to use the original holding and penalty costs as coefficients
of the squared physical inventory and squared backlog, respectively, to obtain an
effective Quadratic Programming (QP) formulation.
Spitter (2005) uses the same experimental setting as de Kok & Fransoo (2003).
The general MIME system under consideration consists of 11 items and is depicted
in figure 8. The 11 item product structure consists of 4 end-items. All end-items con-
tain common component 11. End-items 1 and 2 share component 9, while end-items
3 and 4 share component 10. And each end-item contains a specific component.
We assume that the demand for the end-items is stationary. More precisely, de-
mand for end-item i in consecutive periods is i.i.d. We also assume that the demand
processes for different end-items are uncorrelated. The mean demand is 100 for all
end items. We vary the squared coefficient of variation cv2i for each end-item i as
0.25, 0.5, 1 and 2. The costs structure is as follows
hi = h f = 100 inventory costs end-items, i=1,2,3,4
hi = hs = 10 inventory costs specific components, i=5,6,7,8
hi = hsc = 30 inventory costs semi-common components, i=9,10
hi = hc = 50 inventory costs common components, i=11
For the planned lead times we have analogously to the costs structure the follow-
ing variables

Lk = L f nominal lead time end-items, k=1,2,3,4

Contents 53

/ 5 ?/4 1 /

/ 9

/ 6 4*/ 2 /

/ 11

/ 7 /4* 3 /

/ 10

/ 8 */ 4 /

Fig. 8 Schematic representation of 11 item model.

Li = Ls nominal lead time specific components, i=5,6,7,8

Li = Lsc nominal lead time semi-common components, i=9,10
Li = Lc nominal lead time common components, i=11
We vary the planned lead times (Ls ,Lsc ,Lc ) as follows (1,2,4), (4,2,1) and (1,4,2).
The safety stocks are chosen such that we obtain a non-stockout probability of 95%
for each of the control policies. Here we use the safety stock adjustment procedure
developed in Køhler-Gudum & De Kok (2002), which exploits the translation prop-
erty that is the cornerstone of 1. The results of the comparison between the SBS
concept and the rolling schedule policies are given in table 8, where we present
the inventory costs of each policy and the relative difference between the cost of a
rolling schedule policy and the SBS policy (indicated as ∆. ).
Our experiment shows that the QP rolling scheduling policy is superior to both
LP policies and comes close to the performance of SBS policies, especially for the
cases where the specific components have the longest lead time, i.e. where SBS con-
trols each end-item’s supply chain in isolation. The insertion of the linear allocation
constraints makes the LP formulation more robust. But we may conclude that the
54 Contents

Supply chain inventory cost

cv2i (Ls , Lsc , Lc ) SBS LPst ∆LP LPalloc ∆alloc QP ∆QP
0.25 (1,2,4) 71682 79477 10.9% 73458 2.5% 73332 2.3%
0.25 (4,2,1) 76476 78133 2.2% 78853 3.1% 76765 0.4%
0.25 (1,4,2) 73550 80620 9.6% 74197 0.9% 74172 0.8%
0.5 (1,2,4) 104448 115227 10.3% 106923 2.4% 106679 2.1%
0.5 (4,2,1) 112316 114659 2.1% 115696 3.0% 113381 0.9%
0.5 (1,4,2) 107616 115386 7.2% 108376 0.7% 108362 0.7%
1 (1,2,4) 152203 168533 10.7% 158101 3.9% 157752 3.6%
1 (4,2,1) 165328 169122 2.3% 170590 3.2% 167180 1.1%
1 (1,4,2) 157034 169030 7.6% 159588 1.6% 159722 1.7%
2 (1,2,4) 218551 247578 13.3% 233044 6.6% 232392 6.3%
2 (4,2,1) 245998 249849 1.6% 250075 1.7% 247705 0.7%
2 (1,4,2) 228789 247571 8.2% 235761 3.0% 235694 3.0%

Table 8 Inventory costs of different SCOP functions.

allocation before synchronization concept underlying the SBS policies yields the
best performance in all cases.
Limitations of the above experiment are two-fold:
• Drawing conclusions from the analysis of a single supply chain structure may
not be justified. We developed the 11-item case to represent the key elements
of a general MIME system.In De Kok (2001) a real-world case is reported that
confirms the findings, but clearly further research is needed.
• We have not been able to optimize the control parameters for items i ∈ / E un-
der the rolling schedule policies. As we had to use discrete event simulations
with a run length of 25000 periods or more to ensure accurate point estimates,
where running each case took at least 15 minutes, we considered optimization
of control parameters for items i ∈/ E prohibitive. Given the currently available
CPU’s, it should be possible now to set up an experiment to optimize the control
parameters for all items under rolling schedule policies.
Below further evidence is provided for the effectiveness of SBS policies in com-
parison to other methods to determine control policies for general MIME systems.
We also discuss our empirical findings with the application of SBS policies in real-
life situations.

4.10 Empirical validity of SBS policies

The empirical validity of modelling MIME inventory systems assuming SBS poli-
cies has been extensively studied in a series of MSc theses (Camp (2002), Janssen
(2004), Roose (2007), Bisschop (2007), Uquillas Andrade (2010), Hernandez Wesche
(2012), Van Wanrooij (2012), Radstok (2013), Van Pelt (2015), and Van Cruchten
(2016)). Though it is unknown what policies are used in practice (cf. Section 2), we
Contents 55

found that the calculated fill rate or ready rate under SBS policies was close to the
actual measured ones in several distinct case studies. As inputs we used historical
demand data to calculate mean and variance of end-item demand, historical inven-
tory data to calculate average inventories for all items, historical lot size data to
calculate the average lot size for each item, BoM data and nominal lead time data.
In the software used we converted average lot sizes to nested power-of-2 review
periods. Though this yields a heuristic analysis of the MIME systems considered,
the favorable outcomes suggest a similar policy independent behaviour of MIME
systems as we found for SISE systems (cf. the results in Table 3).
We note here that the nominal lead times should not include safety times for child
items being late. The multi-echelon model explicitly takes into account such delays.
In many cases we found that the main reason for lateness of orders is late starts due
to lack of child item availability. We argued above that under MRP I logic orders are
released without formal check on child item availability. Fortunately, in most MRP
systems nominal lead times are distinguished from safety lead times. The fact that
we found empirical evidence for the validity of multi-echelon models under SBS
policies shows that nominal lead times can be adhered to in practice with a high
probability.
When testing the empirical validity of the SBS policies we identified a major
distinction between MIME systems and SISE systems. In the latter systems, as-
suming some inventory control policy, there is a 1-1 relationship between average
inventory and customer service. In fact, as just mentioned, it seems that there is a
1-1 relationship between average inventory and customer service, irrespective of the
inventory control policy. An higher average inventory of an item yields a higher cus-
tomer service level. This is no longer the case for MIME systems, but in a different
sense than may be expected. Our empirical validation does suggest that average in-
ventories and order frequencies of all items determine end-items’ customer service.
However, we found that increasing average inventories of items does not always
yield higher end-item service levels. The SBS policies provide a formal argument
for this phenomenon.
We mentioned in section 4.5.1 that in convergent MIME systems dead stocks
may emerge as a consequence of not respecting the synchronization inequalities
(20). Dead stocks are defined as the part of an item inventory that does not add to
end-item customer service levels. As under SBS policies we also synchronize (after
allocation), dead stocks cannot emerge under SBS policies. This implies that if an
SBS policy yields a higher average inventory for some item, while all other average
item inventories remain the same, than for at least one end-item customer service
level increases, while all other end-item service levels are not reduced. When start-
ing from actual average inventories, we found that dead stocks naturally emerge. To
provide further understanding we need to define the inventory levels in the decision
node structures, from which we can determine the inventory levels in the underlying
general MIME system.
56 Contents

Variable Definition
E
XC C stationary physical stock in decision node determined by compo-
nent set C and end-item set EC
X̄iact actual average physical stock of item i
Xidead dead stock of item i

With the above definition we can write the average physical stock of an item as
the sum of the average physical stock in decision nodes of which the component set
contains item i.

E
E[Xi ] = ∑ E[XC C ], ∀i. (29)
{(C,EC )|i∈C}

But as in real-life dead stocks may emerge as a consequence of violation of equa-

tions (20), we can write the following equations from which we must derive the
average physical stocks in each decision node.

E
X̄iact = ∑ E[XC C ] + Xidead , ∀i. (30)
{(C,EC )|i∈C}

As illustrated in the case example in section 4.6 in general MIME systems items
are part of component sets in multiple decision nodes. This implies that the set of
linear equations is underdetermined, i.e we have more variables than equations. To
resolve this we need an objective function. An obvious one would be the system’s
customer service level and maximize this. This would yield a nonlinear objective
function and linear constraints, which may lead to prohibitive computation times.
Another approach could be to minimize the value of the dead stocks, which implies
solving an LP. Based on our finding on inventory stock positioning, discussed below
in section 4.11, we developed a simple greedy heuristic aimed at creating as much as
possible inventory downstream in the decision node structures, while respecting the
linear constraints. Once the average physical stocks of the decision nodes are known,
it is rather straightforward to determine the base stock levels that yield these physical
stocks. Given the base stock levels end-item customer service can be calculated. We
concluded that it is easier to find optimal policies for general MIME systems under
SBS policies than to compute the performance of a general MIME system from
historical data.
It should be noted that empirical validation of stochastic models is not trivial.
We already explained that it is important to identify the appropriate performance
indicators for customer service in section 2 for comparison against model fill rate
or model ready rate. We also discussed the impact of high demand volatility in sec-
tion 3.2.3. In a particular case in a high volume supply chain we found that without
removing the end-items with coefficient of variation greater than 1.5 we estimated
an aggregate ready rate of 80%, while the actual aggregate ready rate was 97%. Af-
ter removing the high coefficient of variation end-items, we estimated an aggregate
Contents 57

ready rate of 96.9%, while the actual aggregate ready rate remained 97%. Clearly,
aggregation over end-item ready rates is adding apples and oranges, but estima-
tion of individual end-item aggregate ready rates is similar to throwing dices a few
times. The underlying idea is that applying a consistent modelling approach causes
random variations around the actual targets that cancel out after aggregation. De-
veloping some theoretical background for this statement is highly relevant. In the
above-mentioned MSc projects at various companies with various types of supply
chains, including process industry, FMCG, pharma, and high-tech we observed that
the aggregate model service levels were close to the aggregate real-life service lev-
els.
A generic question concerning empirical validity is how many cases are needed
to support empirical validity. We may argue that given the complexity of the multi-
dimensional functions involved in computing long-run average service levels and
inventory on-hand, it would be very coincidental if the empirical outcomes are in
line with the model outcomes, if the model is incorrect. It would be valuable is
formal statistical methods can be applied to ensure sufficient rigour.
As has been demonstrated in de Kok et al. (2005), SBS policies can be used as the
basis for real-life planning systems that enable to generate material-feasible order
releases in large-scale MIME systems within a split second. To-date, SBS policies
are the only policies that have this capability. Clearly there have been many other
approaches proposed for MIME inventory systems, such as the bounded-demand
model (a.k.a. the guaranteed service model, GSM) proposed by Graves & Willems
(2000) and the stochastic service model (SSM) proposed by Ettl et al. (2000), but
neither of these approaches provide an explicit allocation mechanism that can be
used to ensure material-feasible order releases. The GSM and SSM models have
been developed to compute safety stocks, not to provide a control policy that can
be executed in real-life situations. Below we discuss the consequences of different
modelling assumptions for the same MIME inventory system in more detail.

4.11 Positioning inventory in the supply chain

One of the main contributions of the optimization of MIME inventory systems has
been the insight into the optimal positioning of on-hand inventory capital in the
supply chain subject to end-item service level constraints. The extensive simula-
tion study in Whybark & Yang (1996) is conclusive: minimizing total inventory
in the supply chain implies putting 90% or more of the on-hand inventory at end-
item level. When minimizing inventory capital across the supply chain subject to
end-item service level constraints, the numerical study in de Kok & Fransoo (2003)
shows that by far the most part of the supply chain buffers in time and quantity
should be concentrated at end-item levels. Other inventory capital is allocated to
child items that have parent items that add substantial marginal costs, and in par-
ticular when these child items have long lead times. In most of the supply chains
studied in the MSc projects mentioned in De Kok (2015), the optimal policies do
58 Contents

not hold any stock for many items, implying that the optimal policies create a flow
from purchase items until the end-items. Though flow is strongly advocated by ”lean
thinkers”, we found that most supply chain professionals consider our findings coun-
terintuitive.
These findings are based on assuming SBS policies for inventory control in
MIME inventory systems. As mentioned above the GSM and SSM modelling ap-
proaches are based on different assumptions:
• Bounded demand assumption
The bounded demand assumption, also known as the guaranteed service as-
sumption, was originally coined by Graves & Willems (2000) and their paper
along with Magnanti et al. (2006) has lead to a substantial number of papers
that allow to deal with general structures and various other modelling consid-
erations. The bounded demand assumption allows for setting installation base
stock levels that guarantee all demand is satisfied. The positioning of inven-
tory is optimized by setting so-called service times. An item’s outbound service
time is the time its parent items have to wait for replenishment in addition to
the nominal lead time for shipment or production. An item’s inbound service
time is the additional delay the item itself experiences on top of its nominal
lead time. For every choice of the service times the installation base stock level
is chosen to cover for demand over the sum of nominal lead time and inbound
service time minus the outbound service time.
• Stochastic service assumption
In Graves & Willems (2003), the stochastic service assumption is described as
the assumption that orders are released without a guarantee that they are de-
livered within the nominal lead time, i.e., after a constant delay. In particular,
Graves & Willems (2003) discuss the paper by Ettl et al. (2000) that assumes
installation base stock policies and FCFS as allocation policy. Under this as-
sumption, an approximate analysis is possible.
Note that the main assumption underlying the analysis and optimization of SBS
policies is the balance assumption.
The three approximations can be applied to the same systems, in particular as-
suming constant delays and assuming bounded demand is derived from some per-
centile of the normal or gamma distribution. In Graves & Willems (2003), the
bounded demand assumption (GSM) model is compared with the stochastic ser-
vice (SSM) model for two stylized cases, each derived from a real-world case. We
added the solution from assuming SBS policies to this comparison. We present the
results of the extended comparison in tables 9 and 10.
The results in tables 9 and 10 show the impact of the modelling assumptions.
Not only do we see different total supply chain costs, but more importantly, the al-
location of costs across the supply chain is quite different. This reveals that different
strategic decisions are proposed under different modelling assumptions. In our view
it is important that further research provides deeper insight into the cause of these
differences. It also shows the importance of empirical validation of the models to
assess whether a modelling assumption is justified. This implies a strong need for
Contents 59

Method Safety stock costs

Local DC’s Central DC Packaging Manufacturer Components Total
GSM 347,080 383,567 91,827 0 30,524 852,998
SSM 369,664 399,192 43,557 54,467 60,217 927,097
SBS 493,387 3,069 64,301 5,201 43,442 609,399
Table 9 Cost division across battery supply chain

Method Safety stock costs

Final assembly Main Modules Submodules Other Total
GSM 607,969 0 0 24,751 632,720
SSM 299,472 187,922 129,613 104,870 721,877
SBS 364,734 104,472 55,403 67,086 591,695
Table 10 Cost division across bulldozer supply chain

• more comparisons of different model paradigms on an identical set of bench-

mark problems not necessarily meeting all technical assumptions of the models
and
• more field studies where a user specifies what he prefers to see as a result of the
application of inventory management models, e.g. optimal solutions, robustness
of solutions.

5 Capacitated inventory systems

So far we assumed infinite resource capacity. In this section we discuss the impli-
cations of taking into account finite capacity resources when modelling and ana-
lyzing MIME systems. In subsection 5.1 we discuss the feasibility constraints to
be respected when developing control policies for capacitated MIME systems. In
subsection 5.2 we briefly discuss our findings when comparing alternative rolling
schedule policies. In subsection 5.3 we discuss the main findings from literature on
capacitated serial MIME systems. As currently no results on optimal policies for ca-
pacitated MIME systems other than for serial systems, we discuss ways to implicitly
model finite capacity through nominal lead times.
Taking into account finite resource availability when releasing orders to produc-
tion or transportation is by no means obvious for three reasons:
• Resources process items some time after the order release decision is taken, and
in between events occur that are unknown at the moment of order release
• Resources are mostly used to process more than one item, and the overall pro-
cessing rate depends on the mix of items to be processed.
60 Contents

• Processing rates can be modified when needed, i.e. resources are flexible, but
when what kind of resource flexibility is needed is not known at the moment of
order release.
The vast production planning literature may suggest that resources can be mod-
elled easily in the form of constraints on number of units processed per time unit,
but we should be aware that most models proposed are deterministic: a planning
instance is solved. In a real-life situation the immediate order releases from the so-
lution are implemented and some time later a new planning instance is solved, which
concerns for a large part the same period of time, e.g. a month or a year. In between
the planning instance solutions a myriad of initially unknown events have realized
themselves different from the assumptions made in the model. This makes it un-
clear what objective is appropriate when solving subsequent deterministic instances.
In section 4.9 we discussed the finding that under demand uncertainty SBS policies
outperform LP-based rolling schedule policies for uncapacitated MIME systems (cf.
de Kok & Fransoo (2003). Unfortunately, it is not clear how to extend the concept
of SBS policies for finite resource capacity. To date, we do not know the optimal
policy for capacitated serial MIME systems, unless we impose specific conditions
on the capacity constraints. To our knowledge no results are available for divergent
capacitated MIME systems.
This section starts with the formulation of necessary conditions for feasible order
release decisions, extending inequalities (15) to capacitated MIME systems. Note
that we again adopt the modelling approach that we assume nominal lead times for
each item. These nominal lead times represent the flow time of released orders to
the shop floor as a consequence of stochastic demand, stochastic processing times
and finite capacity. In our view the finite capacity cannot substitute for the nominal
lead times, as is often claimed by authors with a deterministic point of view. Flow
times are not endogenous to the models formulated, but endogenous to the vastly
more complex reality, and thereby exogenous to the models formulated. We discuss
the findings from a numerical study in Spitter (2005). We conclude this section with
a discussion of recent results for serial capacitated MIME systems, which can be
seen as building blocks for finding effective policies for general capacitated MIME
systems.

5.1 Feasibility of order release quantities

Let us first add the variables that determine the resource structure to take into ac-
count. Similar as in de Kok & Fransoo (2003) and Spitter (2005) we assume that
each item is processed at a single resource type.
When deriving order release decisions from a quantitative model that is an ab-
straction of a capacitated MIME inventory system, within the model it should re-
spect the following constraints:
• material availability
Contents 61

Table 11 Defining a multi-echelon inventory system operational characteristics

Variable Definition
Cm,t Amount of capacity available in units of time of resource m in period t
ci Time required to process one unit of item i
Km set of items that can be processed on resource m

• resource availability
• due date targets based on the nominal lead times

Let us assume we are at time 0 and want to create a feasible production plan
respecting the above constraints. Using the notation from Tables 5 and 11 we for-
mulate the following necessary constraints.

N
Gi (t − 1,t] = ∑ ai j r j (t) ≤ Xi (t), ∀i, ∀t ≥ 0 (31)
i=1
t t+1
∑ ri (s) ≥ ∑ Pi (s), ∀i, ∀t ≥ 0 (32)
s=0 s=1
t t+Li
∑ ri (m) ≤ ∑ Pi (s), ∀i, ∀t ≥ 0 (33)
m=0 s=1

∑ ci Pj (t) ≤ Cm,t) , ∀m, ∀t ≥ 1 (34)

i∈Km

Inequality (35) states that the amount of item i needed to release the orders of
its parent items cannot exceed item i’s availability. Inequality (32) states that the
cumulative number of items i received in stock until (and including) time t +s cannot
exceed the cumulative number of item i released until time t + s. Inequality (33)
states that the cumulative number of item i released until time t must have been
produced before time t + Li . Under FIFO this ensures that the orders satisfy their
due dates based on the nominal lead times. Finally, equation (34) states that the total
production time spent on resource type m in period t cannot exceed the available
capacity resource type m. Under these conditions we can apply the nominal lead
time assumption likewise in the uncapacitated case.

Qi (t + Li ] = ri (t), ∀i (35)
62 Contents

5.2 Comparison of rolling scheduling concepts

Building on the promising results of the Quadratic Programming formulation in

a rolling scheduling setting for general MIME systems, Spitter (2005) designed an
experiment similar to the one discussed in section 4.9 with a W-structured BOM and
various assignments of items to resources. On average the QP formulation approach
outperformed the LP formulation approach by 25% lower costs. For robustness of
the QP solution it appeared to be important to choose appropriate weights of the
quadratic physical inventories and backlogs. Choosing these weights consistent with
the allocation fractions defined in equation (26) ensured that in all cases the QP
formulation outperformed the LP formulation. Interestingly, under high resource
utilization scenarios, LP performed very well. This may be due to the fact that under
high utilization long delays for customer orders are inevitable. As nominal lead
times are respected for each item, and nominal lead times are much shorter than
the customer lead times under high utilization, there is perfect knowledge about
customer demand. In that case the allocation problem to be solved is deterministic
and linear by nature, given the linear holding and penalty costs.
There is no benchmark available to judge the quality of the QP rolling schedule
formulation for general MIME systems. Existing results for serial MIME systems
cannot be used, as there is no allocation problem to be solved: each stage in the
serial system has its own resource. As mentioned above, increasing computer power
may allow for solving for small problems, like the W-system, the value iteration
scheme that comes with an SDP formulation of the problem.

5.3 Optimal policies for serial MIME systems

Before discussing results for serial capacitated MIME systems, it is appropriate to

mention that for SISE systems without fixed costs is has been shown in Federgruen
& Zipkin (1986a) and Federgruen & Zipkin (1986b) that the optimal policy is a
modified base stock policy. I.e. each time the system orders, it orders up to a base
stock level, unless the capacity constraint does not allow for that. In that case the
order equals the capacity available expressed in units. Given this result for SISE
systems, it is not difficult to show that for both serial and divergent systems where
only the most upstream stage is capacitated, the optimal policy is a modified echelon
base stock policy for the most upstream stage and a base stock policy for all other
stages.
In their seminal paper on two-stage capacitated serial MIME systems Parker &
Kapuscinski (2004) characterize the optimal policy under the assumption that the
most downstream stage has the tightest capacity. Again they find a modified ech-
elon base stock policy. But they also show that if the capacity is most tight at the
upstream stage, then a modified echelon base stock policy is no longer optimal.
They extend their results to N-echelon serial MIME systems by assuming that the
two most upstream stages are capacitated, the most upstream stage has more ca-
Contents 63

pacity than the one but most downstream stage and all other stages have infinite
capacity.
Janakiraman & Muckstadt (2009) extend the results of Parker & Kapuscinski
(2004) to the case of N-echelon capacitated MIME systems where all stages have
the same capacity. Assuming integer demand they find that so-called multitier base
stock policies are optimal: each stage has multiple base stock levels that are targeted
dependent on the state of the system. This extends the finding of Parker & Kapus-
cinski (2004) who mentioned a hidden base stock level in case the most upstream
stage had the tightest capacity. This multidimensionality of the optimal policy sug-
gests that it cannot be expected that computationally attractive optimal policies can
be found for general capacitated MIME systems. Remember that even for uncapac-
itated divergent systems we need the balance assumption to ensure echelon base
stock policies are optimal that can efficiently be computed recursively.
Recently Huh & Janakiraman (2010), Huh et al. (2010), and Huh et al. (2016)
explore the properties of serial capacitated MIME systems under base stock policies.
The results seem to be promising, as under high end-item service level regimes base
stock policies are asymptotically optimal and can be calculated straightforwardly.
Such asymptotic arguments may be extended to more complicated MIME systems.
This seems to be a promising route to go.
In light of this recent work it is important to mention the contributions of Glasser-
man & Tayur (1994) and Glasserman & Tayur (1996), who use Infinitesimal Per-
turbation Analysis (IPA) to find optimal base stock policies for capacitated serial
MIME systems. This simulation-based approach seems to be a promising candidate
to find optimal base stock policies for general MIME systems, as IPA is a generic
method to exploit cost gradient information from discrete event simulation sam-
ple paths. On the other hand, the fact that M/M/1 queues need at least 3 million
customer arrivals before simulation results converge to long-run results from math-
ematical analysis, seems to suggest that capacitated MIME systems need long-run
times before ensuring that results are representative. One may (rightfully) argue that
such a number of jobs or orders is not realistic, but then we should be aware that
short-run simulation outcomes are very sensitive to the coincidental inputs regarding
customer demands and item processing times. There seems to be a methodological
issue here: what to compare simulation outcomes with, if comparing against long-
run optimal policies is impossible.

5.4 Implicit modelling of finite capacity

Given the unsolved problem of finding close-to-optimal policies for general MIME
systems, while in reality we are faced with finite resource availability, we may want
to follow a heuristic route. In our extensive discussion on human behaviour and re-
source flexibility in section 2 we provided arguments for modelling MIME systems
with nominal lead times. Our empirical findings reported in section 4.10 suggest
that such modelling yields valid results. That opens the route for decomposition of
64 Contents

the problem of capturing the impact of finite capacity by determining a nominal

lead time for each resource or resource-item combination that can be realized with
a high probability. An estimate for such a nominal lead time could be derived from
queueing models, assuming that the average throughput time from the model is a
good estimate of the average flow time of a production order (cf. our discussion in
section 2). Though this approach is feasible, given the knowledge on queuing mod-
els for production departments (cf. Hopp & Spearman (2011)), it will not be easy
to test in simulation environments. This relates to our arguments in section 2 that in
practice there is always more flexibility than we model for. Given the state of the
art of research on general capacitated MIME systems, it is clearly worthwhile to
explore this alternative, albeit less rigourous, perspective.

6 Conclusion

In this paper we discussed various aspects of inventory modelling and analysis. In-
ventory modelling is about abstracting from real-world details, maintaining only
the essence of the inventory management process. This essence is about ordering
in time in the right quantity at the right frequency. However, in practice a myriad
of feed-back mechanisms and responses are available that have a substantial impact
on the performance of inventory management. Because this array of interventions
and their impact cannot be captured in mathematical models, we propose to dis-
tinguish between inventory management performance before (human) interventions
that deviate from the ones assumed in the model and the inventory performance
after interventions. This leads to the concept of Intervention Independent Perfor-
mance (IIP) and Intervention Dependent Performance (IDP). The IIP indicators can
be used to derive inventory control parameters, the IDP indicators measure the real
performance. Our assumption is that we can identify a relationship between IIP per-
formance and IIP performance. We report the feasibility of this approach implicitly,
when discussing the empirical validity of inventory models for both SISE systems
and MIME systems. This empirical validity is based on consultancy projects exe-
cuted by the authors and many MSc thesis projects executed by IE students from
Eindhoven University of Technology, of which for some references are provided.
Over almost 6 decades the objective of inventory modelling has been minimiza-
tion of discounted or long-run average costs. The underlying idea is that we should
identify the relevant costs to take into account. However, on the stock market (ex-
pected) Return On Investment is the objective. Using the simplest possible inventory
model, the EOQ model of Harris (1913), we discuss the findings in Trietsch (1995),
who showed that under the ROI maximization objective optimal order quantities de-
pend differently on model parameters than under the cost minimization objective. In
short, the classical EOQ is always greater than the ROI-minimizing order quantity
R∗ . In fact, EOQ equals R∗ in a limiting sense, as fixed investments go to infinity.
Probably the most important result discussed is that average inventory and aver-
age order frequency determine the end-item customer service levels in both SISE
Contents 65

and MIME inventory systems. This is of prime importance for future research. Bar-
ing in mind that optimal policies cannot be found for general MIME systems, it
suffices to develop results for mathematically tractable policies that provide full
freedom in setting average inventories and average order frequencies for individual
items. In that case we can exploit generic optimization techniques to find optimal
policy parameters, and possibly exploit specific properties of the relevant objective
function to find fast optimization algorithms.
A potential building block for such fast optimization techniques is the generic
validity of the Newsvendor fractile under optimal policies within a class of policies.
The net stock translation property holds for most known SISE inventory control
policies. Combination of this result with the Safety Stock Adjustment Procedure of
Køhler-Gudum & De Kok (2002) enables efficient simulation-based optimization
of the end-item control policy parameters in general MIME systems. Open question
is of this result can be extended to optimization of control policy parameters of
non-end-items.
The extension of this result holds under Synchronized Base Stock policies. We
show that SBS policies can control general MIME systems under the assumption
of periodic review and nominal lead times. SBS policies reveal the hierarchical
item order release decision structure embedded in general MIME systems. Simula-
tion experiments provide evidence of the cost-effectiveness of SBS policies against
mathematical-programming-based rolling schedule policies. Close-to-optimal SBS
policies can be determined efficiently by recursively solving one-dimensional gen-
eralized Newsvendor equations. Case studies show that modelling real-life supply
chains with the SBS policy framework yields empirically valid results.
Major challenges regarding the analysis and optimization of control policies of
general MIME systems remain. We mention here the need to develop results for
MIME systems with fixed setup and ordering costs that naturally lead to lot sizing
constraints. And our short discussion on capacitated MIME systems makes clear
that, even though base stock policies seem to be effective for serial MIME systems,
it is yet unclear how to analyze general capacitated MIME systems under base stock
policies. Presumably, there is another Gordian knot to be cut.

7 Acknowledgements

I am indebted to Henk Tijms who taught me to appreciate development of accurate

approximations instead of making assumptions for the convenience of tractability,
which could rule out application of results in practice.
I am indebted to Ed Silver, for his book with Rein Peterson that combines rigour
and deep insights into what inventory management (and many more subjects in Op-
erations Management) in practice is about. The 1985 edition is my inventory man-
agement Bible. But even more am I thankful to Ed for his open mind, when I started
sending him letters, questioning some modelling assumptions in this edition.
66 Contents

And I am indebted to Will Bertrand, who provided the conceptual production and
inventory control framework within which I could embed my work on multi-echelon
inventory systems and hierarchical planning to ensure that the models I worked on
could be implemented in practice and at least would provide relevant insights to
practitioners.

References

Axsäter, S. (2003), ‘Supply chain operations: Serial and distribution inventory systems’, Hand-
books in operations research and management science 11, 525–559.
Axsäter, S. (2015), Inventory Control, International Series in Operations Research & Management
Science, Springer International Publishing.
URL: https://books.google.nl/books?id=v9YjCgAAQBAJ
Bisschop, J. (2007), Supply chain performance evaluation : application of the synchronised base
stock policy in a high-tech complex equipment supply chain with contract manufacturers,
Master’s thesis, School of Industrial Engineering, Eindhoven University of Technology.
URL: https://pure.tue.nl/ws/files/47041546/631639-1.pdf
Camp, B. (2002), Startrek supply chain planning : modeling, optimization and generalization, Mas-
ter’s thesis, School of Industrial Engineering, Eindhoven University of Technology.
URL: https://pure.tue.nl/ws/files/46793392/561053-1.pdf
Clark, A. J. & Scarf, H. (1960), ‘Optimal policies for a multi-echelon inventory problem’, Man-
agement science 6(4), 475–490.
Daganzo, C. F. (2005), Logistics systems analysis, Springer.
De Kok, A. (1991a), Basics of inventory management, Technical Report 510, 521-525, Tilburg
University.
De Kok, A. (1991b), Basics of inventory management (part 5): The (r,b,q)-model, Technical report.
De Kok, A. (1993), Demand management in a multi-stage distribution chain, Technical Report
93-35, Eindhoven University of Technology.
De Kok, A. (1998), Inventory control with manufacturing lead time flexibility, Technical Report
345, Eindhoven University of Technology.
De Kok, A. (2001), Comparison of supply chain planning concepts for general multi-item, multi-
echelon systems, Technical report, Research Report TUE/TM/LBS/01-03. Eindhoven:
Technische Universiteit Eindhoven.
URL: https://pure.tue.nl/ws/files/3577889/552034.pdf
De Kok, T. (2015), Buffering against uncertainty in high-tech supply chains, in ‘Proceedings of the
2015 Winter Simulation Conference’, IEEE Press, pp. 2991–3000.
De Kok, T. G. (2003), ‘Ruin probabilities with compounding assets for discrete time finite hori-
zon problems, independent period claim sizes and general premium structure’, Insurance:
Mathematics and Economics 33(3), 645–658.
de Kok, T. G. (2017), ‘Modelling short-term manufacturing flexibility by human intervention and
its impact on performance’, accepted for publication in International Journal of Production
Research .
de Kok, T. G. & Fransoo, J. C. (2003), Planning supply chain operations: Definition and com-
parison of planning concepts, in ‘Supply Chain Management: Design, Coordination and
Operation’, Vol. 11 of Handbooks in Operations Research and Management Science, Else-
vier, pp. 597 – 675.
URL: http://www.sciencedirect.com/science/article/pii/S0927050703110122
Contents 67

De Kok, T. G. & Visschers, J. W. (1999), ‘Analysis of assembly systems with service level con-
straints’, International Journal of Production Economics 59(1), 313–326.
de Kok, T., Janssen, F., Van Doremalen, J., Van Wachem, E., Clerkx, M. & Peeters, W. (2005),
‘Philips electronics synchronizes its supply chain to end the bullwhip effect’, Interfaces
35(1), 37–48.
Dekker, R., Kleijn, M. & De Kok, A. (1998), The break quantity rule’s effect on inventory costs in
a 1-warehouse, n-retailers distribution system, Technical Report 7.
Diks, E. B. & De Kok, A. (1998), ‘Optimal control of a divergent multi-echelon inventory system’,
European journal of operational research 111(1), 75–97.
Diks, E. & De Kok, A. (1999), ‘Computational results for the control of a divergent n-echelon
inventory system’, International Journal of Production Economics 59(1), 327–336.
Doğru, M. K., De Kok, A. & Van Houtum, G. (2009), ‘A numerical study on the effect of the
balance assumption in one-warehouse multi-retailer inventory systems’, Flexible services
and manufacturing journal 21(3-4), 114–147.
Edgeworth, F. (1888), ‘The mathematical theory of banking’, Journal of the Royal Statistical So-
ciety 51, 113–127.
Eppen, G. & Schrage, L. (1981), ‘Centralized ordering policies in a multi-warehouse system with
lead times and random demand’, Multi-level production/inventory control systems: Theory
and practice 16, 51–67.
Ettl, M., Feigin, G. E., Lin, G. Y. & Yao, D. D. (2000), ‘A supply network model with base-stock
control and service requirements’, Operations Research 48(2), 216–232.
Federgruen, A. & Zipkin, P. (1984), ‘Computational issues in an infinite-horizon, multiechelon
inventory model’, Operations Research 32(4), 818–836.
Federgruen, A. & Zipkin, P. (1986a), ‘An inventory model with limited production capacity
and uncertain demands i. the average-cost criterion’, Mathematics of Operations Research
11(2), 193–207.
Federgruen, A. & Zipkin, P. (1986b), ‘An inventory model with limited production capacity and
uncertain demands ii. the discounted-cost criterion’, Mathematics of Operations Research
11(2), 208–215.
Fogarty, D. & Hoffmann, T. (1983), Production and Inventory management, South-Western Pub-
lishing Co., Cincinnati, Ohio.
Glasserman, P. & Tayur, S. (1994), ‘The stability of a capacitated, multi-echelon production-
inventory system under a base-stock policy’, Operations Research 42(5), 913–925.
Glasserman, P. & Tayur, S. (1996), ‘A simple approximation for a multistage capacitated
production-inventory system’, Naval Research Logistics 43(1), 41–58.
Graves, S. C. & Willems, S. P. (2000), ‘Optimizing strategic safety stock placement in supply
chains’, Manufacturing & Service Operations Management 2(1), 68–83.
Graves, S. C. & Willems, S. P. (2003), ‘Supply chain design: safety stock placement and supply
chain configuration’, Handbooks in operations research and management science 11, 95–
132.
Harris, F. (1913), ‘How many parts to make at once’, Factory, The Magazine of Management
10(2), 135–136.
Hernandez Wesche, E. (2012), Impacts op implementing a retailer cross-dock on the western eu-
rope procter & gamble supply chain, Master’s thesis, School of Industrial Engineering,
Eindhoven University of Technology.
URL: https://pure.tue.nl/ws/files/46910223/739767-1.pdf
Hopp, W. J. & Spearman, M. L. (2011), Factory physics, Waveland Press.
Huh, W. T. & Janakiraman, G. (2010), ‘Base-stock policies in capacitated assembly systems: Con-
vexity properties’, Naval Research Logistics (NRL) 57(2), 109–118.
68 Contents

Huh, W. T., Janakiraman, G. & Nagarajan, M. (2010), ‘Capacitated serial inventory systems: sam-
ple path and stability properties under base-stock policies’, Operations Research 58(4-part-
1), 1017–1022.
Huh, W. T., Janakiraman, G. & Nagarajan, M. (2016), ‘Capacitated multiechelon inventory sys-
tems: Policies and bounds’, Manufacturing & Service Operations Management 18(4), 570–
584.
Janakiraman, G. & Muckstadt, J. A. (2009), ‘A decomposition approach for a class of capacitated
serial systems’, Operations Research 57(6), 1384–1393.
Janssen, F. (2004), Voorraadverlaging door scm bij diosynth, Master’s thesis, School of Industrial
Engineering, Eindhoven University of Technology.
URL: https://pure.tue.nl/ws/files/46832256/577518-1.pdf
Karaarslan, A., Kiesmüller, G. & De Kok, A. (2013), ‘Analysis of an assemble-to-order system with
different review periods’, International Journal of Production Economics 143(2), 335–341.
Kiesmüller, G. P., de Kok, T. G., Smits, S. R. & van Laarhoven, P. J. (2004), ‘Evaluation
of divergent n-echelon (s, nq)-policies under compound renewal demand’, OR Spectrum
26(4), 547–577.
Kilger, C., Meyr, H. & Stadtler, H. (2015), ‘Supply chain management and advanced planning:
concepts, models, software, and case studies’.
Kleinrock, L. (1965), ‘A conservation law for a wide class of queueing disciplines’, Naval Research
Logistics Quarterly 12(2), 181–192.
Køhler-Gudum, C. & De Kok, A. (2002), A safety stock adjustment procedure to enable target
service levels in simulation of generic inventory systems, Technical report.
URL: https://pure.tue.nl/ws/files/3596094/554365.pdf
Magnanti, T. L., Shen, Z.-J. M., Shu, J., Simchi-Levi, D. & Teo, C.-P. (2006), ‘Inventory placement
in acyclic supply chain networks’, Operations Research Letters 34(2), 228–238.
Nahmias, S. & Olsen, T. L. (2015), Production and operations analysis, Waveland Press.
Orlicky, J. (1975), Material requirements planning.
Parker, R. P. & Kapuscinski, R. (2004), ‘Optimal policies for a capacitated two-echelon inventory
system’, Operations Research 52(5), 739–755.
Radstok, K. (2013), Fast & slow freight distribution in the fast moving consumer goods industry,
Master’s thesis, School of Industrial Engineering, Eindhoven University of Technology.
URL: https://pure.tue.nl/ws/files/46922783/754756-1.pdf
Roose, S. (2007), Rethinking inbound operations management at procter & gamble mechelen :
multi-echelon inventory management applied in process industry, Master’s thesis, School
of Industrial Engineering, Eindhoven University of Technology.
URL: https://pure.tue.nl/ws/files/47021144/627272-1.pdf
Rosling, K. (1989), ‘Optimal inventory policies for assembly systems under random demands’,
Operations Research 37(4), 565–579.
Silver, E. A., Pyke, D. F. & Thomas, D. J. (2016), Inventory and Production Management in Supply
Chains, CRC Press.
Silver, E., Pyke, D. & Peterson, R. (1998), Inventory Management and Production Planning and
Scheduling, Wiley, New York.
Song, J.-S. & Zipkin, P. (2003), ‘Supply chain operations: Assemble-to-order systems’, Handbooks
in operations research and management science 11, 561–596.
Spitter, J. M. (2005), Rolling schedule approaches for supply chain operations planning, Technis-
che Universiteit Eindhoven Eindhoven.
URL: https://pure.tue.nl/ws/files/2089632/200511140.pdf
Trietsch, D. (1995), ‘Revisiting roq: Eoq for company-wide roi maximization’, The Journal of the
Operational Research Society 46, 507–515.
Contents 69

Uquillas Andrade, R. (2010), An integral supply chain operations planning system for a global
pharmaceutical company, Master’s thesis, School of Industrial Engineering, Eindhoven
University of Technology.
URL: https://pure.tue.nl/ws/files/46966985/668984-1.pdf
Van Cruchten, A. (2016), Multi-echelon safety stock optimization under supply, process and de-
mand uncertainties as a part of operational risk management, Master’s thesis, School of
Industrial Engineering, Eindhoven University of Technology.
URL: https://pure.tue.nl/ws/files/46933553/845023-1.pdf
Van der Heijden, M. (1997), ‘Supply rationing in multi-echelon divergent systems’, European
Journal of Operational Research 101(3), 532–549.
Van der Heijden, M., Diks, E. & De Kok, A. (1997), ‘Stock allocation in general multi-echelon
distribution systems with (r, s) order-up-to-policies’, International Journal of Production
Economics 49(2), 157–174.
van Houtum, G.-J., Scheller-Wolf, A. & Yi, J. (2007), ‘Optimal control of serial inventory systems
with fixed replenishment intervals’, Operations Research 55(4), 674–687.
Van Pelt, T. (2015), Multi-echelon inventory management at sligro food group n.v., Master’s thesis,
School of Industrial Engineering, Eindhoven University of Technology.
URL: https://pure.tue.nl/ws/files/47009988/789280-1.pdf
Van Wanrooij, M. (2012), Strategic supply chain planning in a multi-echelon environment : identi-
fication of the codp location constrained by controllability and service requirements, Mas-
ter’s thesis, School of Industrial Engineering, Eindhoven University of Technology.
URL: https://pure.tue.nl/ws/files/46910380/739787-1.pdf
Vollmann, T., Berry, W., Whybark, D. & Jacobs, F. (2005), Manufacturing Planning and Control
for Supply Chain Management.
Whitin, T. M. (1953), The Theory of Inventory Management, Princeton.
Whybark, D. C. & Yang, S. (1996), ‘Positioning inventory in distribution systems’, International
Journal of Production Economics 45(1-3), 271–278.
Willems, S. P. (2008), ‘Data setreal-world multiechelon supply chains used for inventory optimiza-
tion’, Manufacturing & Service Operations Management 10(1), 19–23.
Wilson, J. M. (2016), ‘The origin of material requirements planning in frederick w. taylors planning
office’, International Journal of Production Research 54(5), 1535–1553.
Zipkin, P. (2000), Foundations of Inventory Management, McGraw-Hill, Boston.