See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/225861808

Forecasting for Inventory Management of Service Parts

Chapter · January 2008

DOI: 10.1007/978-1-84800-011-7_20

CITATIONS READS

40 6,334

4 authors, including:

John Boylan
Buckinghamshire New University
93 PUBLICATIONS   3,021 CITATIONS   

SEE PROFILE

All content following this page was uploaded by John Boylan on 26 May 2014.

The user has requested enhancement of the downloaded file.

 Comment [MSOffice1]:
Chapter 20. To appear in 2007:
In (eds: Kobbacy, K.A.H. and
20 Murthy, D.N.P.) Complex System
Maintenance Handbook, Springer.

Forecasting for Inventory Management

of Service Parts

John E. Boylan and Aris A. Syntetos

20.1 Introduction
Service parts are ubiquitous in modern societies. Their need arises whenever a
component fails or requires replacement. In some sectors, such as the aerospace
and automotive industries, a very wide range of service parts are held in stock, with
significant implications for availability and inventory holding. Their management
is therefore an important task.
A distinction should be drawn between preventive maintenance and corrective
maintenance. Demand arising from preventive maintenance is scheduled and is
deterministic, at least in principle. Demand arising from corrective maintenance,
after a failure has occurred, is stochastic and requires forecasting.
Fortuin and Martin (1999) categorise the contexts for service logistics as
follows:
• Technical systems under client control (eg machines in production
departments, transport vehicles in a warehouse);
• Technical systems sold to customers (eg telephone exchange systems,
medical systems in hospitals);
• End products used by customers (eg TV sets, personal computers, motor
cars)
In the first context, there is usually a specialist department within the client
organization performing maintenance activities and managing service parts
inventories. In the second context, a specialist department within the vendor
organization will generally undertake these tasks. In both cases, a large amount of
information is known by the vendor, or can be shared with the vendor. This
information may include scheduled (preventive) maintenance activities, times
between failures, usage rates and condition of equipment.
When a wealth of data is available, it is possible to identify explanatory
variables which may be used to predict the demand of service parts. For example,
Ghobbar and Friend (2002) showed that the average demand interval for aircraft
2 J.E. Boylan and A.A. Syntetos

spare parts depends on the aircraft utilization rate, the component overhaul life and
the type of primary maintenance process. In a further study, Ghobbar and Friend
(2003) showed how forecast accuracy depends on various characteristics of the
demand process, including the seasonal period length, as well as the primary
maintenance process. Hua et al (2006) used two zero-one explanatory variables,
plant overhaul and equipment overhaul, to help predict demand of spare parts in
the petrochemical industry. In other cases, explanatory variables have been used to
predict part of the demand for a Stock Keeping Unit (SKU). For example,
Kalchschmidt et al (2006) identified clusters of customers whose sales were
correlated with promotional activities and clusters of customers that were
unaffected, using appropriate forecasting methods for each group.
In the third context, parts are used by consumers and much less information is
available. Fortuin and Martin (1999:957) commented, “Clients are anonymous,
their usage of consumer products and their ‘maintenance concept’ are not known”.
Most demand arises from purely corrective maintenance (eg on TV sets, personal
computers) required in the case of a defect. Even when preventive maintenance
occurs (eg on motor cars), prediction is complicated by the ‘maintenance concept’
of consumers being unknown. For example, customers may not bring in their cars
at the correct time for a service, or may not bring them in at all. In many practical
situations where end products are used by consumers, the vendor must gauge
demand for service parts from the demand history alone. Such demand patterns are
often sporadic, with occasional ‘spikes’ of demand. Alternatively, demand for an
SKU may be decomposed into regular and irregular components (Kalchschmidt et
al, 2006). In both cases, sporadic demand for service parts poses a considerable
challenge to those responsible for managing inventories. It is this challenge that
will be addressed in this chapter.
The remainder of the chapter is structured as follows: In the next section we
address issues pertinent to the classification of service parts for forecasting and
inventory management related purposes. Parametric and non-parametric
approaches to forecasting service parts requirements are then discussed in section 3
and 4 respectively. In section 5, we present various metrics appropriate for
measuring the performance of the inventory management system whereas in
section 6 we review the limited number of studies that provide empirical evidence
on: i) the performance of forecasting methods for service parts and ii) the empirical
fit of statistical distributions to the corresponding underlying demand patterns.
Finally, the conclusions of our work are summarized in section 7.

20.2 Classification of Service Parts

Service parts for consumer products are highly varied, with differing costs, service
requirements and demand patterns. Classification of Stock Keeping Units (SKUs)
is widely adopted by organizations, but the method of classification varies widely.
This is to be expected, as classification serves a number of different purposes.
 Forecasting and Inventory Management for Service Parts 3

20.2.1 Service Requirement

The first aim of classification is to determine service requirements. It is common

for organizations to segment their service parts, assigning higher service-level
targets to some segments than others. A direct approach is to classify according to
a part’s service criticality. The ‘criticality’ may be determined informally or by
formal methods such as Failure Mode, Effects and Criticality Analysis (FMECA).
According to this method, criticality analysis is defined as “A procedure by which
each potential failure mode is ranked according to the combined influence of
severity and probability of occurrence” (Department of Defense, 1980:3). This
approach is likely to be more appropriate for those situations where technical
systems are being managed by the vendor or are under client control. However, it
may also be applicable to service parts for consumers. For example, safety-critical
automotive components, such as brakes, may be assigned to a higher criticality
category than automotive accessories, such as furry dice.
Alternatively, an ABC (Pareto) classification can be used to determine service
requirements. A Pareto report lists all SKUs in descending order, by total volume,
or total value of sales. An ABC analysis by value is often used as a proxy for
criticality, with the A items being assumed to be the most critical and requiring the
highest service levels. Some authors also argue that the sophistication of the
replenishment method should reflect the ABC classification: “For a true C item
the low total of replenishment, carrying and shortage costs implies that, regardless
of the type of control system used, we cannot achieve a sizable absolute savings in
these costs. Therefore, the guiding principle should be to use simple procedures
that keep the control costs per SKU quite low…” (Silver, Pyke and Peterson, 1998:
359). For a single SKU, this is undoubtedly correct. However, for many hundreds
or thousands of SKUs, the argument has less force. The additional savings accruing
from more sophisticated forecasting and stock control methods potentially
outweigh any additional investment or system running costs.
A further disadvantage of the ABC classification is that it is not obvious in
what different ways the categories should be treated. (This basic requirement of
inventory classification schemes was first discussed by Williams, 1984.) In
particular, the choice of forecasting methods for slower demand items depends on
the degree of intermittence and the variability of demand, neither of which are fully
captured by a Pareto analysis by volume or value. Therefore, if a Pareto
classification is adopted, it may be advantageous to supplement it with further
categorizations. For example, classification based on value of sales is often used to
determine the frequency of orders to be placed, whereas classification of demand
characteristics (to be examined later in this section) is a more effective way to
determine the order levels and the safety stocks.
Categorization of service parts by cost is common practice. In some
organizations, a two-way classification by cost and volume is employed. This
allows greater flexibility in adjusting service targets, by category, in order to
achieve overall targets at minimum cost. It is a slightly more sophisticated
variation of the ABC approach, again requiring supplementary classifications for
forecasting purposes.
4 J.E. Boylan and A.A. Syntetos

20.2.2 Inventory Decision

A Product Life Cycle approach is often used in marketing, with three phases of
growth, maturity and decline. A similar classification may be adopted for stock
control, with the phases aligned directly to the decisions required for the inventory
management of service parts.
Fortuin (1980) suggested three phases: initial, normal and final. In the initial
phase, when the part is introduced, there are two decisions: i) should the item be
stocked and ii) if so, what are the initial stock requirements? In the normal phase,
an inventory policy must be determined and the parameters estimated. If an Order-
Up-To (OUT) policy is adopted, for example, then the Order-Up-To-Level must be
calculated. As the part nears the end of its life, suppliers may become reluctant to
manufacture small volumes, as required by clients, particularly if the part has high
manufacturing set-up costs. In this final phase, a decision must be taken on the size
of a single order to cover all remaining demand (sometimes known as an ‘all time
buy’). Teunter (1998) analysed this problem from a theoretical perspective, while
Teunter and Fortuin (1998) reported a case-study of a company facing such a
decision.

20.2.3 Forecasting Appraoch

Forecasting approaches may be broadly divided into two categories:

• Dependent on explanatory variables (causal methods);
• Dependent only on the history of demand (time-series methods)
A classification of service parts according to the Product Life Cycle can assist
in choosing the better approach. As discussed in the first section of this chapter, the
choice of forecasting approach is mainly determined by the availability of data on
explanatory variables, such as the timing of preventive maintenance activities.
However, the forecasting approach is also driven by the availability of demand
history data which, in turn, is determined by the stage of the service part’s life
cycle.
Causal methods are particularly useful in the initial phase, when the part is
introduced, since the lack of an adequate length of demand history precludes the
use of extrapolative time-series methods. Models linking sales to promotional
expenditure, for example, can be applied. In the normal phase, which is the focus
of this chapter, causal methods are used when maintenance activities are under the
control of the vendor or the client (if the client is not an end-consumer). For
consumer clients, historical data for the explanatory variables are usually not
available, and time-series methods are used to forecast service parts’ requirements.
In the final phase, when an ‘all time buy’ from a supplier is required, extrapolative
methods can be applied. For example, a regression model on the logarithm of sales
against time may be used, assuming an exponential decline in demand over time.
 Forecasting and Inventory Management for Service Parts 5

20.2.4 Forecasting Method

Faster moving service parts are commonly forecast using time-series methods. The
specific method that should be employed depends on the characteristics of the
demand pattern. For non-intermittent demand, exponential smoothing methods are
often used, with appropriate variants for trended, damped trended and seasonal
data. For intermittent demand, with some periods showing no demand at all,
different methods are needed.
Demand is said to be ‘intermittent’ if it is “infrequent in the sense that the
average time between consecutive transactions is considerably larger than the unit
time period, the latter being the interval of forecast updating” (Silver, Pyke and
Peterson, 1998:127). An item with ‘erratic demand’ is “one having primarily small
demand transactions with occasional very large transactions” (Silver, 1970:87).
Intermittent and erratic demand patterns are very common amongst service
parts. If an item is both intermittent and erratic, it is said to be ‘lumpy’. The
following graph shows examples of intermittent and lumpy demand patterns (based
on annual demand history for two service parts used in the aerospace industry).

70

60

50
Demand (Units)

40

30

20

10

0
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Time period
Slow demand Lumpy demand

Fig. 20.1. Intermittent and lumpy demand patterns

A further approach to the categorization of service parts is to examine the sources

of intermittence and erraticness in the demand pattern. Bartezzaghi et al (1996)
identified the following factors contributing to these demand characteristics:

1. Numerousness of potential customers

2. Frequency of customer requests
3. Heterogeneity of customers (measured by Gini’s index)
6 J.E. Boylan and A.A. Syntetos

4. Variety of customers’ requests (measured by the coefficient of variation of

the demand of a single customer)
5. Correlation between customers’ requests.

The first and second factors determine the intermittence of demand. In response
to this intermittence, for those SKUs with very few customers, it may become
feasible to liaise directly with them, and to enhance forecasts accordingly.
The third and fourth factors determine the ‘erraticness’ of demand. As orders
become more irregular, exploiting early information at the customer level becomes
more attractive. Of course, such early indications are not always available. This
will often be the case when addressing consumer demand. It is also possible that
early confirmed orders may give a good indication of final orders. This is
particularly useful when there is a strong correlation between customers’ demands.

Numerousness of customers
Intermittence
Frequency of individual orders

Correlation between customers’ requests Lumpiness

Heterogeneity of customers
Erraticness
Variety of customers’ requests

Fig. 20.2. Categorization based on the sources of demand characteristics

For those items without early indicators, forecasting must be undertaken using a
purely time-series approach. This is usually linked to a demand distribution, so that
inventory levels may be set to achieve high percentage service level targets. Many
inventory management systems make distributional assumptions of demand
according to the ABC classification. For example, A and B items may be taken to
be normally distributed, whilst C items are assumed to be Poisson. In practice,
however, many service parts have demand that is more erratic than Poisson
(sometimes known as ‘over-dispersed’). The Poisson Dispersion Index (ratio of the
variance to the mean of demand, including zero demands) can be used to classify
SKUs as Poisson or non-Poisson. If the index is close to unity, then a Poisson
distribution is indicated; if the index is greater than unity, then other distributions,
such as the Negative Binomial, may be more appropriate, or a non-parametric
approach may be required, as discussed in section 20.4.
 Forecasting and Inventory Management for Service Parts 7

An obvious way to classify service parts is by frequency of demand. As

demand occurrence becomes more infrequent, with some periods having no
demand at all, a number of difficulties emerge. From a forecasting perspective,
methods such as Single Exponential Smoothing (SES) can no longer be
recommended (Croston, 1972). Secondly, assumptions of normality of demand
become unsustainable, as demand becomes more skewed. Johnston and Boylan
(1996) examined the conditions under which Croston’s method (designed for
intermittent demand and reviewed in the next section of this chapter) is more
accurate than Single Exponential Smoothing. The authors concluded, on the basis
of simulation of a wide range of conditions, that Croston’s method is more accurate
(in terms of Mean Square Error) when the average demand interval (p) exceeds
1.25 review periods. In a recent case-study, Boylan et al (2006) showed that
overall forecast accuracy is robust to the choice of break-points above 1.25 but
much less so to values below 1.25. For practical applications, break-points may be
determined using simulation studies. The important point is that it is preferable to
identify conditions for superior forecasting performance, and then to categorise
demand based on these results, rather than the other way round.
A complementary method of classification is by variability of demand size.
From a forecasting perspective, Syntetos et al (2005) re-examined the comparison
between methods such as Croston (1972), based on intermittent demand, and SES.
This study was based on comparison of approximate expressions for the theoretical
Mean Square Error of different methods. The authors identified two key
categorization variables, namely the average demand interval (p) and the squared
2
coefficient of variation of demand size (CV ). (Note that the latter measure ignores
zero demand periods). Comparisons between forecasting methods yield regions of
superior performance. The performance of Single Exponential Smoothing depends
on whether it is assessed at all points in time, or only immediately after demand
occurrences (which trigger stock-orders in most inventory systems). The
performance of Croston’s method does not depend on this timing. Suppose that
SES is compared with Croston’s method, for the forecasts made immediately after
a demand occurrence; then the matrix presented in Figure 20.3 is obtained.

Low High
p=1.34 (break-point)
High
Erratic Lumpy
(Croston) (Croston)
CV2=0.28
Smooth Intermittent
(SES) (Croston)
Low

Fig. 20.3. Categorisation of SKUs by forecast accuracy

8 J.E. Boylan and A.A. Syntetos

In summary, service parts may be in the initial, normal or final phases of their life
cycle. In this chapter, we focus attention on the normal phase. Although service
parts may be classified as A, B or C in a Pareto analysis, it is likely that most parts
will be categorized as C. The service requirements for the part may be guided by
criticality and cost considerations, as well as the ABC classification. Further
refinements are necessary to the Pareto classification in order to allocate the most
appropriate forecasting methods to each SKU.
Enhancement of the ABC classification in the manner described above gives a
coherent approach to classification according to forecasting performance and a
foundation for theoretically-informed usage of terms such as ‘erratic’, as shown in
Figure 20.4 (after Syntetos, 2001, adapted by Boylan et al, 2006).

High
Intermittent
Mean inter-
demand interval
Non-intermittent
Low

Mean demand size Slow

Low
AND/OR

High
Erratic Lumpy
AND
Coefficient of variation
of demand sizes
Non-erratic Clumped
Low AND

Fig. 20.4. Categorisation of demand patterns for service parts

As in Figure 20.2, a ‘lumpy’ SKU is defined as one that is both ‘intermittent’ and
‘erratic’; definitions of ‘slow’ and ‘clumped’ are also included. Figure 20.4 offers a
different perspective from Figure 20.2. The former diagram shows the measures
that may be used to classify SKUs as intermittent and erratic, whereas the latter
showed the factors that lead to intermittence and erraticness. An understanding of
both issues is required for effective forecasting of service parts.
 Forecasting and Inventory Management for Service Parts 9

20.3 Parametric Forecasting

Practical parametric approaches to inventory management rely upon estimates of
some essential demand distribution parameters. The decision parameters of the
inventory systems (such as the Re-Order Point or the Order-Up-To-Level) are then
based on these estimates.
Different inventory systems require different variables to be forecasted. Some
of the most well cited, for example (R, s, S) policies (Naddor, 1975; Ehrhardt and
Mosier, 1984), require only estimates of the mean and variance of demand. (In
such systems, the inventory position is reviewed every R periods and if the stock
level drops to the re-order point s enough is ordered to bring the stock up to the
order-up-to-level S.)
In other cases, and depending on the objectives or constraints imposed on the
system, such estimates are also necessary, although they do not constitute the ‘key’
quantities to be determined. We may consider, for example, an (R, S) or an (s, Q)
policy operating under a fill-rate constraint – known as P2 and discussed in section
20.5. (In the former case, the inventory position is reviewed periodically, every R
periods, and enough is ordered to bring it up to S. In the latter case, there is a
continuous review of the inventory position and as soon as that drops to, or below,
s an order is placed for a fixed quantity Q.) In those cases we wish to ensure that
x% of demand is satisfied directly off-the-shelf and estimates are required for the
probabilities of any demands exceeding S or s. Such probabilities are typically
estimated indirectly, based on the mean demand and variance forecast in
conjunction with a hypothesized demand distribution. Nevertheless, a
reconstruction of the empirical distribution through a bootstrapping (non-
parametric) procedure would render such forecasts redundant and this issue is
further discussed in the following section. Similar comments apply when these
systems operate under a different service driven constraint: there is no more than
x% chance of a stock-out during the replenishment cycle (this service measure is
known as P1). Consequently, we need to estimate the (100-x)th percentile of the
demand distribution.
In summary, parametric approaches to forecasting involve estimates of the
mean and variance of demand. In addition, a demand distribution needs also to be
hypothesized, in the majority of stock control applications, for the purpose of
estimating the quantities of interest. Issues related to the hypothesized demand
distribution are addressed in the following sub-section. The estimation of the mean
and variance of demand is addressed in sub-sections 20.3.2 and 20.3.4 respectively.

20.3.1 The Demand Distribution

Demand for service parts is most commonly intermittent in nature. The demand
pattern is characterized by infrequent demands, often of variable size, occurring at
irregular intervals. Consequently, as discussed in the next sub-section, it is
preferable to model demand from constituent elements, i.e. the demand size and
inter-demand interval. Therefore, compound theoretical distributions (that
explicitly take into account the size-interval combination) are typically used in
such contexts of application. We first discuss some issues related to modelling
10 J.E. Boylan and A.A. Syntetos

demand arrivals and hence inter-demand intervals. We then extend our discussion
to compound demand distributions.
In a service parts demand context, two demand generation processes have
dominated the literature. If time is treated as a discrete (whole number) variable,
demand may be generated based on a Bernoulli process, resulting in a geometric
distribution of the inter-demand intervals. When time is treated as a continuous
variable, the Poisson demand generation process results in negative exponentially
distributed inter-arrival intervals.
There is sound theory in support of both geometric and exponential distribution
for representing the time interval between successive demands. There is also
empirical evidence in support of both distributions (eg Dunsmuir and Snyder,
1989; Kwan, 1991; Willemain et al, 1994; Janssen, 1998; Eaves, 2002). With
Poisson arrivals of demands and an arbitrary distribution of demand sizes, the
resulting distribution of total demand over a fixed lead time is compound Poisson.
Inter-demand intervals following the geometric distribution in conjunction with an
arbitrary distribution for the sizes, results in a compound binomial distribution.
Regarding the compound Poisson distributions, the stuttering Poisson, which is
a combination of a Poisson distribution for demand occurrence and a geometric
distribution for demand size, has received the attention of many researchers (for
example: Gallagher, 1969; Ward, 1978; Watson, 1987). Another possibility is the
combination of a Poisson distribution for demand occurrence and a normal
distribution for demand sizes (Vereecke and Verstraeten, 1994), although the latter
assumption has little empirical support. Quenouille (1949) showed that a Poisson-
Logarithmic process yields a negative binomial distribution (NBD). When order
occasions are assumed to be Poisson distributed and the order size is not fixed but
follows a logarithmic distribution, total demand is then negative binomially
distributed over time.
Another possible distribution for representing demand is the gamma
distribution. The gamma distribution is the continuous analogue of the NBD and
“although not having a priori support [in terms of an explicit underlying
mechanism such as that characterizing compound distributions], the gamma is
related to a distribution which has its own theoretical justification” (Boylan, 1997:
168). The gamma covers a wide range of distribution shapes, it is defined for non-
negative values only and it is generally mathematically tractable in its inventory
control applications (Burgin and Wild, 1967; Burgin, 1975; Johnston, 1980).
Nevertheless if it is assumed that demand is discrete, then the gamma can be only
an approximation to the distribution of demand. At this point it is important to note
that the use of both NBD and gamma distributions requires estimation of the mean
and variance of demand only. In addition, and as discussed in section 20.6, there is
empirical evidence in support of both distributions and therefore they are
recommended for practical applications.
Vereecke and Verstraeten (1994) presented an algorithm developed for the
implementation of a computerised stock control system for spare parts in a
chemical plant. Ninety per cent of the items were classified as lumpy, with the
remaining ten per cent consisting of slow or fast movers. The demand was assumed
to occur as a Poisson process with a package of several pieces being requested at
each demand occurrence. The parameters of the distribution of the demand size can
 Forecasting and Inventory Management for Service Parts 11

be estimated from the variance and the average of the demand history data of each
item. The resulting distribution of demand per period was called a ‘Package
Poisson’ distribution. The same distribution has appeared in the literature under the
name ‘hypothetical SKU’ (h-SKU) Poisson distribution (Williams, 1984), where
demand is treated as if it occurs as a multiple of some constant, or ‘clumped
Poisson’ distribution, for multiple item orders for the same SKU of a fixed ‘clump
size’ (Ritchie and Kingsman, 1985; please also refer to Figure 20.4 where a
definition of ‘clumped’ demand is offered). In an earlier work, Friend (1960) also
discussed the use of a Poisson distribution for demand occurrence, combined with
demands of constant size. The ‘Package Poisson’ distribution requires, as the
Poisson distribution itself, an estimate of the mean demand only.
If demand occurs as a Bernoulli process and orders follow the Logarithmic-
Poisson distribution (which is not the same as the Poisson-Logarithmic process that
yields NBD demand) then the resulting distribution of total demand per period is
the log-zero-Poisson (Kwan, 1991). The log-zero-Poisson is a three parameter
distribution and requires a rather complicated estimation method. Moreover, it was
found by Kwan (1991) to be empirically outperformed by the NBD. Hence, the
log-zero Poisson cannot be recommended for practical applications. One other
compound binomial distribution appeared in the literature is that involving
normally distributed demand sizes (Croston, 1972; 1974). However, and as
discussed above, a normality assumption is unrealistic and therefore the
distribution is not recommended for practical applications.

20.3.2 Estimation of Mean Demand: Size – Interval Methodology

Single Exponential Smoothing (SES) and Simple Moving Averages (SMA) are
often used in practice to forecast intermittent demand. Both methods have been
shown to perform satisfactorily on real service parts data. However, the ‘standard’
forecasting method for such items is considered to be Croston’s method (Croston,
1972, as corrected by Rao, 1973). Croston suggested treating the size of orders
( z t ) and the intervals between them ( p t ) as two separate series and combining
their expenentially weighted moving averages (obtained using SES) to achieve a
forecast of the demand per period. (Recently, some adaptations of Croston’s
method have appeared in the literature that rely upon SMA rather than SES
estimates and such modifications are further discussed later in this sub-section.)
In Croston’s work, both demand sizes and intervals were assumed to have
constant means and variances, for modelling purposes, and demand sizes and
demand intervals to be mutually independent. Demand was assumed to occur as a
Bernoulli process. Subsequently, the inter-demand intervals are geometrically
distributed (with mean p ). The demand sizes were assumed to follow the normal
distribution (with mean μ and variance σ 2 ).
These assumptions have been challenged in respect of their realism (see, for
example, Willemain et al, 1994) and they have also been challenged in respect of
their theoretical consistency with Croston’s forecasting method. The latter issue is
further discussed in sub-section 20.3.3.
12 J.E. Boylan and A.A. Syntetos

Croston’s method works in the following way: SES estimates of the average
size of the demand ( ẑ t ) and the average interval between demand incidences
(p̂ t ), are made after demand occurs (using the same smoothing constant value,
α ). If no demand occurs, the estimates remain exactly the same. The forecast of
demand per period ( Yˆt ) is given by: Yˆt = zˆ t / pˆ t . If demand occurs in every time
period, Croston’s estimator is identical to SES. For constant lead times of length
L , the mean lead-time demand estimate ( YˆL ) is then obtained as follows:

YˆL = LYˆt (20.1)

Despite the theoretical superiority of such an estimator, modest benefits were

recorded in the literature when the method was actually applied on real data.
Syntetos and Boylan (2001) showed that Croston’s estimator is biased. The bias is
introduced by estimating the probability of demand occurrence from the average
inter-demand interval (inversion bias). This is now explained in some more detail.
We start with Croston’s assumptions:

Ε( z t ) = Ε( zˆ t ) = μ (20.2a)
Ε( p t ) = Ε( pˆ t ) = p (20.2b)

According to Croston, the expected estimate of demand per period in that case
would be: Ε(Yˆt ) = Ε( zˆ t / pˆ t ) = Ε( zˆ t ) / Ε( pˆ t ) = μ / p (i.e. the method is
unbiased).
If it is assumed that estimators of demand size and demand interval are
independent, then

⎛ zˆ ⎞ ⎛ 1 ⎞
Ε⎜⎜ t ⎟⎟ = Ε( zˆ t )Ε⎜⎜ ⎟⎟ (20.3)
ˆ
⎝ pt ⎠ ⎝ pˆ t ⎠

but

⎛ 1 ⎞ 1
Ε⎜⎜ ⎟⎟ ≠ (20.4)
ˆ
⎝ t⎠
p Ε ( pˆ t )

and therefore Croston’s method is biased. It is clear that this result does not
depend on Croston’s assumptions of stationarity and geometrically distributed
demand intervals.
More recently, Boylan and Syntetos (2003), Syntetos and Boylan (2005) and
Shale et al (2006) presented correction factors to overcome the bias associated with
Croston’s approach. Some of these papers discuss: i) Croston’s applications under
 Forecasting and Inventory Management for Service Parts 13

a Poisson demand arrival process and ii) estimation of demand sizes and intervals
using an SMA (using the ratio of the former to the latter as an estimate of demand
per period). The correction factors are summarized in the following table (where k
is the length of the moving average and α is the smoothing constant for SES).

Table 20.1. Bias correction factors

Demand generation process
Bernoulli Poisson
α α
1− 1−
SES 2 2 −α
Syntetos and Boylan, 2005 Shale et al, 2006
Estimation
k k −1
SMA k +1 k
Boylan and Syntetos, 2003 Shale et al, 2006

At this point it is important to note that SMA and SES are often treated as
equivalent when the average age of the data in the estimates is the same (Brown,
1963). A relationship links the number of points in an arithmetic average (k) with
the smoothing parameter of SES ( α ) for stationary demand. Hence it may be used
to relate the correction factors presented in Table 20.1 for each of the two demand
generation processes considered. The linking equation is:

2 −α
k= (20.5)
α

20.3.3 Method – Model Inconsistencies

Snyder (2002) pointed out that Croston’s model assumes stationarity of demand
intervals and yet an SES estimator is used, implying a non-stationary demand
process. The same comment applies to demand sizes. Snyder commented that this
renders the model and method inconsistent and he proposed some alternative
models, and suggested a new forecasting approach based on parametric
bootstrapping. Shenstone and Hyndman (2005) developed this work by examining
Snyder’s models. In their paper they commented on the wide prediction intervals
that arise for non-stationary models and recommended that stationary models
should be reconsidered. However, they concluded, “...the possible models
underlying Croston’s and related methods must be non-stationary and defined on a
continuous sample space. For Croston’s original method, the sample space for the
underlying model included negative values. This is inconsistent with reality that
demand is always non-negative” (Shenstone and Hyndman, op. cit: 389-390).
14 J.E. Boylan and A.A. Syntetos

In summary, any potential non-stationary model assumed to be underlying

Croston’s method must have properties that do not match the demand data being
modeled. Obviously, this does not mean that Croston’s method and its variants are
not useful. Such methods do constitute the current state of the art in intermittent
demand parametric forecasting. An interesting line of further research would be to
consider stationary models for intermittent demand forecasting rather than
restricting attention to models implying Croston’s method. For example, Poisson
autoregressive models have been suggested to be potentially useful by Shenstone
and Hyndman (2005).

20.3.4 Estimation of Demand Variance

In parametric forecasting and inventory control applications, estimating the

variability of the lead-time demand forecast error is equally important to estimating
the level of demand itself. In this sub-section we address issues related to the
estimation of the error variance but it is important to note that such an estimate
may not always be required. Under the assumption of Poisson distributed demand,
for example, an estimate of the mean demand only would be sufficient.
In practical applications, and assuming constant lead-times, the variance of the
lead-time forecast error is most often taken as the sum of the error variances of the
individual forecast periods. In particular, if L is the length of the lead-time
(constant), and σˆ t is the standard deviation of the one-step ahead forecast error at
time t , then the standard deviation of the lead-time forecast error ( σˆ L ) is
estimated as follows:

σˆ L = Lσˆ t (20.6)

In theory, the standard deviation of the one-step-ahead forecast error σˆ t can be

estimated by using either the Mean Squared Error (MSE) or the Mean Absolute
Deviation (MAD) approach. However, the ‘smoothing’ versions of those error
measures are most often used in practice to improve the responsiveness of the
system (see, for example, Silver et al, 1998):

σˆ t = MSE t (20.7)

where MSE t = α (Yt −1 − Yˆt −1 ) 2 + (1 − α ) MSE t −1 , or

σˆ t ≈ 1.25 MAD t (20.8)

where MADt = α Yt −1 − Yˆt −1 + (1 − α ) MADt −1

 Forecasting and Inventory Management for Service Parts 15

In the above calculations, α is the smoothing constant, Yt −1 the actual demand

in period t − 1 and Yˆt −1 the forecast of demand for period t − 1 .
If the mean demand level fluctuates over time (steady state model or
Autoregressive Integrated Moving Average – ARIMA process of order (0,1,1)), it
has been shown (Johnston and Harrison, 1986) that equation (20.6) neglects any
correlation between the estimates of demand. This correlation exists, at least in
part, because of the uncertainty in the estimate of the true underlying level of
demand that is carried from one period to another.
When using SES, under the above model formulation, the standard deviation of
the lead time forecast error was shown (Johnston and Harrison, op. cit.) to be
correctly calculated as follows:

σˆ L = L + α ( L − 1) L(1 + α (2 L − 1 / 6)) σˆ t (20.9)

Under the stationary mean model assumption (the demand level is assumed to
be constant) the forecast error correlation still exists because of the uncertainty
associated with the variance of the forecasts, which is carried forward from one
period to another. In addition, if a biased estimator is in place to forecast future
demand requirements, the auto-correlation can be also attributed to the bias. This
issue has been analytically addressed by Strijbosch et al (2000) and Syntetos et al
(2005).

20.4 Non-Parametric Forecasting

As discussed in section 2, demand for service parts may often be lumpy in nature.
Considering Figure 20.3, such SKUs are characterized by very infrequent demand
occurrences (intermittence) coupled with highly erratic demand sizes, when
demand occurs.
Croston’s method and its variants (in conjunction with an appropriate
distribution) have been reported to offer tangible benefits to stockists forecasting
intermittent demand. (Relevant empirical evidence follows in section 20.6.)
Nevertheless, there are certainly some restrictions regarding the degree of
lumpiness that may be dealt with effectively by any parametric distribution. In
addition to the average inter-demand interval, the coefficient of variation of
demand sizes has been shown in the literature to be very important from a
forecasting perspective for demand classification purposes. However, as the data
become more erratic, the true demand size distribution may not comply with any
standard theoretical distribution. This challenges the effectiveness of any
parametric approach. When SKUs exhibit a lumpy demand pattern such as that
presented in Figure 20.1, one could argue that only non-parametric approaches
may provide opportunities for further improvements in this area.
Willemain et al (2004) developed a patented non-parametric forecasting
method for intermittent demand data. Their method is not model-based but instead
is a heuristic that combines a Markov process, bootstrapping and ‘jittering’ to
16 J.E. Boylan and A.A. Syntetos

simulate an entire distribution for lead-time demand rather than a single forecast.
The method works according to the following steps:

1 Obtain historical demand data in chosen time buckets (e.g. days, weeks,
months)
2 Estimate transition probabilities for two-state (zero vs. non-zero) Markov
model
3 Conditional on last observed demand, use Markov model to generate a
sequence of zero/non-zero values over forecast horizon
4 Replace every non-zero state marker with a numerical value sampled at
random, with replacement, from the set of observed non-zero demands
5 ‘Jitter’ the non-zero demand values – this is effectively an ad-hoc
procedure designed to allow greater variation than that already observed.
The process enables the sampling of demand size values that have not been
observed in the demand history
6 Sum the forecast values over the horizon to get one predicted value of
Lead Time Demand (LTD)
7 Repeat steps 3 – 6 many times
8 Sort and use the resulting distribution of LTD values.

Willemain, Smart and Schwarz (2004: 381) argued that “… we need to assess
the quality not of a point forecast of the mean but of a forecast of the entire
distribution”, but they conceded that it is impossible to compare this on an item-
specific basis. Instead, the authors recommended pooling percentile estimators
across items and measuring the conformance of the observations (expressed using
the corresponding percentiles) to a uniform distribution. The researchers claimed
significant improvements in forecasting accuracy achieved by using their approach
over Single Exponential Smoothing and Croston’s method. (Issues related to
assessing forecasting performance are further considered in the next section.)
Gardner and Koehler (2005) criticized this study in terms of its methodological
arrangements and experimental structure, pointing out that:
• Willemain et al did not use the correct lead time demand distribution for
either SES or Croston’s method. This was a two-fold criticism consisting of
arguments against the use of equation (20.6) for estimating the lead-time
demand variance (please refer to sub-section 20.3.4) and the use of the
normal distribution for representing demand;
• They did not consider published modifications to Croston’s method such as
the estimator proposed by Syntetos and Boylan (2005)
Further empirical evidence is required in order to develop our understanding of
the benefits offered by such a non-parametric approach. In particular, a comparison
between the recently developed adaptations of Croston’s method – see Table 20.1
(in conjunction with an appropriate distribution) with the bootstrapping approach
should prove to be beneficial from both theoretical and practitioner perspectives.
 Forecasting and Inventory Management for Service Parts 17

20.5 Performance Measurement

In assessing the performance of an inventory management system, there are two
essential measures: stock-holding cost and service level. Stock-holding costs are
relatively straightforward to interpret. They are generally calculated as a
percentage of the value of inventory investment, where the percentage takes into
account such factors as the cost of capital, insurance, warehousing and
obsolescence costs. Often, the same percentage is applied to all service parts.
However, one could argue that slow-moving service parts should attract a higher
percentage cost, since these parts are at the highest risk of obsolescence.
‘Service level’ is generally interpreted as ‘off the shelf availability’ but the way
in which it is measured varies. Three common measures are defined as follows
(Silver et al, 1998):
• The fraction of replenishment cycles in which total demand can be
delivered from stock (known as P1). This is equivalent to a specified
probability of no stock-outs during a replenishment cycle;
• The fraction of total demand that can be delivered from stock (known as
P2). This is also called the ‘fill rate’;
• The fraction of time during which there is stock on the shelf (known as P3).
This is sometimes used when equipment is needed for emergency purposes
The ‘fill rate’ is probably the measure with the greatest appeal to practitioners,
since it relates most directly to customer satisfaction. Care is needed with its
application, since different results are obtained if it is calculated over a lead-time or
over all time. If unsatisfied demand is back-ordered, Brown (1967) showed that:

LS
P2 = 1 − (1 − P2LT ) (20.10)
Q

LT
where P2 is the measure over lead-time, P2 is the measure over all time, L is the
lead-time, S is total demand in a year, Q is the order-quantity and it is assumed that
LS < Q.
Ronen (1982) showed that, if unsatisfied demand is lost, then:

1
P2 = (20.11)
LS
(1 − P2LT ) + 1
Q

These measures are based on the fraction of units satisfied from stock. Some
organizations also use measures that relate to the successful completion of an
‘order-line’ for a number of units of the same SKU. Typically, these are based on
the fraction of order-lines completely satisfied (partial satisfaction does not count).
Boylan and Johnston (1994) identified relationships between such measures and
fill-rates.
18 J.E. Boylan and A.A. Syntetos

In addition to these standard measures, other suggestions have been made.

Gardner (1990) recommended the use of trade-off curves, showing the effect of
inventory investment on the average delay in filling backorders. Separate curves
are drawn for each forecasting method, allowing the manager to see at a glance if
one method dominates the others.
Sani and Kingsman (1997) proposed the use of average regret measures. The
service regret is the amount each method falls short of the maximum service level
over all methods for that SKU. (The ‘method’ may be a forecasting or inventory
method). The regret is then divided by the maximum service level and the ratios
are averaged across all SKUs. A cost regret measure is defined similarly. This
approach allows more detailed assessments of the interaction between forecasting
and inventory methods. Eaves and Kingsman (2004) suggested assessment of
forecast performance according to implied stock-holdings. These are based on a
calculation of the exact safety margin providing a maximum stock-out of zero. The
advantage of this approach is that it gives monetary values of stock-savings.
However, these savings may not be achieved in practice using a standard stock
control method based on the mean and variance of lead-time forecasts.
Whilst it is essential to assess stock-holding costs and service levels, it is also
important to be able to diagnose the reasons for any deterioration in these
measures. Boylan and Syntetos (2006) argue that, since this may arise as a result of
forecasting methods or inventory rules (see Figure 20.5 below), the accuracy of
forecasting methods should also be monitored.

Forecasting Stock-holding
Stock
Method Costs
Management

Inventory System Service

Rules Level

Fig. 20.5. Inventory system performance measurement

Forecast error measures may be used to detect changes in forecast accuracy over
time or to determine the relative accuracy of two (or more) forecasting methods.
For faster moving service parts, a measure that serves both purposes and is easy to
interpret is the Mean Absolute Percentage Error (MAPE):

1 n Y − Yˆt
MAPE = ∑
n t =1
100 t
Yt
(20.12)

where n is the number of histrorical forecasts included in the error measure, Yt is

the observation at time t and Yˆt is the forecast of demand at time t.
 Forecasting and Inventory Management for Service Parts 19

Unfortunately, this error measure is not defined for zero observations. This
rules out its application for service parts with intermittent demand. An alternative
measure, suggested by Makridakis (1993) and used by Makridakis and Hibon
(2000), is the Symmetric MAPE (sMAPE):

1 n | Yt − Yˆt |
sMAPE = ∑ | Y + Yˆ | / 2
n t =1
100 (20.13)
t t

However, as pointed out by Syntetos (2001) and by Boylan and Syntetos

(2006), the sMAPE will always be 200 for any period when the actual demand is
zero, regardless of the size of the error. Therefore, the sMAPE does not
discriminate between forecast methods in this case. Since it does not allow a
satisfactory comparison of forecast methods, it is inappropriate for intermittent
demand.
A simple measure that can be used to assess the bias of forecasting methods is
the Mean Error (ME):

1 n
ME = ∑ (Yt − Yˆt )
n t =1
(20.14)

This measure is simple to interpret: if it is close to zero, then the forecast

method is unbiased; negative values indicate that forecasts are consistently too
high, while positive values show that forecasts are too low. Its use is recommended
for intermittent service parts.
If a forecast method has high forecast errors, but is approximately unbiased,
then the positive and negative errors cancel one another out, yielding a Mean Error
close to zero. To capture the degree of error, regardless of sign, other error
measures are required. The Mean Absolute Error (MAE) is often used for an
individual SKU and is defined as follows:

1 n
MAE = ∑ Yt − Yˆt
n t =1
(20.15)

This error measure should not be averaged over a whole set of parts, since it
may be dominated by a few SKUs with large errors. To avoid this problem, four
alternatives have been suggested: the MAE: Mean ratio, the Geometric Mean
Absolute Error, the Percentage Better measure and the Mean Absolute Scaled
Error. Each of these measures will be reviewed in turn.
Hoover (2006) proposed the application of the MAE:Mean ratio for intermittent
demand:
20 J.E. Boylan and A.A. Syntetos

∑Y t − Yˆt n
t =1
n
∑Y t − Yˆt
MAE : Mean = n
= t =1
n
(20.16)
∑Y
t =1
t ∑Y
t =1
t

n
This measure is robust to outlying data and is easy to interpret. Hyndman
(2006) observed that the MAE:Mean ratio assumes that the data is stable over time
and that, for seasonal intermittent data, the measure may become unreliable. This
problem may be overcome, for non-trended data, by calculating the measure over a
full set of seasonal cycles. If the data is trended, however, then Hyndman’s
criticism stands. Therefore, the MAE:Mean ratio can be recommended for non-
trended intermittent service parts.
A second alternative to the MAE is the Geometric Mean Absolute Error
(GMAE) defined below, for a single series:

1/ n
⎛ n ⎞
GMAE = ⎜⎜ ∏ Yt − Yˆt ⎟⎟ (20.17)
⎝ i =1 ⎠

This can be generalized across series by taking the Geometric Mean again to
obtain the Geometric Mean (across series) of the Geometric Mean (across time) of
the Absolute Errors (GMGMAE):

1/ N
⎛ N ⎛ n ⎞
1/ n
⎞
GMGMAE = ⎜ ∏ ⎜⎜ ∏ Yit − Yˆit ⎟⎟ ⎟ (20.18)
⎜ i =1 ⎝ t =1 ⎠ ⎟
⎝ ⎠

where Yit is the observation for the ith SKU at time t, Yˆit is the forecast of
demand for the ith SKU at time t, and N is the number of SKUs.
An outlying observation, producing a large error by any statistical method, will
affect the GMAE similarly for all methods, and so the ratio of the GMAE for one
method to another will be robust to outliers. (The same robustness property applies
to the GMGMAE). This was first shown by Fildes (1992), using a general
argument, and applied to intermittent data by Syntetos and Boylan (2005). In fact,
these authors used a slightly more complex measure, the Geometric Root Mean
Square Error (GRMSE); however, Hyndman (2006) pointed out that the GRMSE
and the GMAE are identical.
Although the measure is robust to outliers, it is sensitive to zero errors (Boylan
and Syntetos, 2006). Just one exact forecast will yield a zero error and a zero
GMAE, regardless of the size of the other errors. This problem may be overcome,
 Forecasting and Inventory Management for Service Parts 21

for stationary errors, by using the Geometric Mean (across series) of the Arithmetic
Mean (across time) of the Absolute Errors (GMAMAE):

1/ N
⎛ N ⎛1 ni
⎞⎞
GMAMAE = ⎜⎜ ∏ ⎜⎜ ∑ Yit − Yˆit ⎟⎟ ⎟⎟ (20.19)
⎝ i =1 ⎝ ni t =1 ⎠⎠

This measure collapses to zero only if a series has zero forecast errors for all
periods of time, and so is more robust to zero errors than the GMGMAE. The
measure is also robust to occasional large forecast errors, provided the remaining
errors are stable, and are not unduly affected by trend or seasonality. It can
therefore be recommended for application in these cases.
Another approach, which is simple to use and interpret, is the Percentage
Better method. According to this approach, for each service part, one forecast
method is compared to another according to a criterion such as Mean Error or
Geometric Root Mean Square Error (Syntetos and Boylan, 2005). The Percentage
Better shows the percentage of series for which one method has the lower error.
This approach is robust to large forecast errors and the results can be subjected to
formal statistical tests (Syntetos, 2001). It is a useful measure, although it does not
quantify the degree of improvement in forecast error.
Hyndman (2006) recently suggested a new error measure for intermittent
demand. This measure, known as the Mean Absolute Scaled Error (MASE), is
defined as follows:

MASE = mean( qt ) (20.20a)

Yt − Yˆt
qt = (20.20b)
1 n
∑ Yt − Yt −1
n − 1 i =2

The errors are scaled based on the in-sample MAE from the naïve forecasting
method (i.e. the forecast for the next period is this period’s observation). The
measure is robust to outliers, and is valid for all non-constant series.
Hyndman (2006) gave an example of the application of the MASE on
intermittent data from a major Australian lubricant manufacturer. He compared the
out-of-sample MASE of four methods: naïve, overall mean, Single Exponential
Smoothing and Croston’s method. The naïve method has the lowest MASE. This
result is valid statistically, but is counter-intuitive from an inventory-management
perspective. Boylan and Syntetos (2006) commented that the naïve method is
sensitive to large demands and will generate high forecasts. Its use will almost
certainly lead to over-stocking and possibly to obsolescence. This example
highlights the danger of relying on statistical error measures alone. As noted earlier
in this section, attention should always be paid to the stock-holding and service
implications of different forecasting methods. Improvements in forecasting
accuracy do not necessarily translate into improved stock-control performance.
22 J.E. Boylan and A.A. Syntetos

However, if stock-control performance has deteriorated, then forecast error

measures can be used to diagnose problems with forecasting methods, and to
suggest alternatives.

20.6 Empirical Evidence

Empirical evidence on the performance of forecasting methods for service parts is
not extensive. The same is true regarding the empirical fit of various distributions
to the underlying demand patterns of such SKUs. In this section the relevant
studies are reviewed.

20.6.1 Statistical Distributions

Kwan (1991) conducted research to identify the theoretical distributions that best
fit the empirical distributions of demand sizes, inter-demand intervals and demand
per unit time period for low demand items. Regarding inter-demand intervals, both
the geometric and the negative exponential distribution were found to provide a
good fit to the demand patterns observed. The geometric distribution was also
found to be a reasonable approximation to the distribution of inter-demand
intervals, for real demand data, by Dunsmuir and Snyder (1989) and Willemain et
al (1994). Janssen (1998) tested the Bernoulli demand generation process on a set
of empirical data obtained from a Dutch wholesaler of fasteners. The results
indicated that the Bernoulli demand generation process is a reasonable
approximation for intermittent demand processes. Finally, Eaves (2002) examined
the demand patterns associated with 6,795 service parts from the Royal Air Force
(UK). The findngs of this detailed study provide support for both Poisson and
Bernoulli processes. In particular, the geometric distribution was found to provide
a statistically significant fit (5% significance level) to 91% of his sample whereas
the negative exponential distibution fitted 88% of the demand histories examined.
Kwan (1991) tested the empirical fit of the log-zero-Poisson (lzP) and negative
binomial (NBD), amongst other possible underlying demand distributions. The
NBD was found to be the best, fitting 90% of the SKUs. Boylan (1997) tested the
goodness-of-fit of four demand distributions (NBD, lzP, Condensed Negative
Binomial Distribution (CNBD) and gamma distribution) on real demand data. The
CNBD arises if we consider a condensed Poisson incidence distribution
(‘censored’ Poisson process in which only every second event is recorded)
assuming that the mean rate of demand incidence is not constant, but varies
according to a gamma distribution. The empirical sample used for testing
goodness-of-fit contained the six months histories of 230 SKUs, demand being
recorded weekly. The analysis showed strong support for the NBD. The results for
the gamma distribution were also encouraging, although not as good, for slow
moving SKUs, as the NBD.
 Forecasting and Inventory Management for Service Parts 23

20.6.2 Demand Estimators

Willemain et al (1994) compared SES and Croston’s method on both theoretically

generated and empirical intermittent demand data (54 series). They concluded that
Croston's method is robustly superior to exponential smoothing and can provide
tangible benefits to stockists dealing with intermittent demand. A very important
feature of their research, though, was the fact that industrial results showed very
modest benefits as compared with the simulation results.
Sani and Kingsman (1997) compared the performance (service level and
inventory costs) of various empirical and theoretically proposed stock control
policies for low demand items as well as that of various forecasting methods
(SMA, SES, Croston) on 30 service parts. Their results indicated i) the very good
overall performance of SMA; ii) the fact that stock control policies that have been
developed in conjunction with specific distributional assumptions – such as the
Power Approximation (explicitly built upon the assumption of a compound
Poisson underlying demand pattern – please also refer to section 20.3) perform
particularly well.
Willemain et al (2004) assessed the forecast accuracy of SES, Croston’s
method (both in conjunction with a hypothesised normal distribution) and the non-
parametric approach that they proposed (please refer to section 20.4) on 28,000
service inventory items. They concluded that the bootstrap method was the most
accurate forecasting method and that Croston’s method had no significant
advantage over SES. As discussed in section 20.4, some reservations have been
expressed regarding the study’s methodology. Nevertheless, the bootstrapping
approach is intuitively appealing for very lumpy demand items. More empirical
studies are needed to substantiate its forecast accuracy in comparison with other
methods.
Syntetos and Boylan (2005) conducted an empirical investigation to compare
the forecast accuracy of SMA, SES, Croston’s method and a bias-corrected
adaptation of Croston’s estimator (termed the Syntetos-Boylan Approximation,
SBA, please refer to Table 20.1). The forecast accuracy of these methods was
tested, using a wide range of forecast accuracy metrics, on 3,000 service parts from
the automotive industry. The results demonstrated quite conclusively the superior
forecasting performance of the SBA method. In a later project, Syntetos and
Boylan (2006) assessed the empirical stock control implications of the same
estimators on the same 3,000 SKUs. The results demonstrated that the increased
forecast accuracy achieved by using the SBA method (also known as the
‘Approximation Method’) is translated to a better stock control performance
(service level achieved and stock volume differences). A similar finding was
reported in an earlier research project conducted by Eaves and Kingsman (2004).
They compared the empirical stock control performance (implied stock holdings
given a specified service level) of the above discussed estimators on 18,750 service
parts from the Royal Air Force (UK). They concluded that ‘the best forecasting
method for a spare parts inventory is deemed to be the approximation method’
(Eaves and Kingsman, 2004: 436).
24 J.E. Boylan and A.A. Syntetos

20.7 Conclusions
Service parts, particularly those subject to corrective maintenance, present a
considerable challenge for both forecasting and inventory management. If stocking
decisions are made injudiciously, then the result will be poor service or excessive
stock-holdings, possibly leading to obsolescence. Conversely, effective forecasting
and stock control will lead to cost savings and improved customer service
A number of stock-control methods may be employed for slow-moving service
parts. Sani and Kingsman (1997) recommended the (R, s, S) policy, based on its
inventory cost and service performance in an empirical study. However, empirical
evidence is not extensive, and further research is needed in this area.
Classification of service parts is an essential element in their management. Four
purposes are served by classification:
• Determination of service targets;
• Establishment of inventory decisions;
• Choice of forecasting approach;
• Choice of forecasting method
Determining service level requirements may be supported by a criticality
classification, undertaken using management judgment or a more formal approach,
such as an assessment of the risk and severity of a part failure. Alternatively, a
Pareto classification may be used as a proxy for criticality, with the A items being
deemed the most important. A variation on this approach is to use a matrix of cost
and sales volume to determine service requirements.
Inventory decisions relate directly to a Product Life Cycle classification. Initial
provisioning decisions must be taken during the initial phase of the life cycle.
Decisions should be taken regarding stocking locations, too. In the normal phase,
the inventory policy and the appropriate parameters must be determined. The
inventory rules depend on forecasts of demand over lead-time, so the most
appropriate forecasting method should be chosen. In the final phase, an ‘all time
buy’ requires a decision on the final order quantity.
The Product Life Cycle can also be used to help determine the forecasting
approach: causal or time-series. Causal methods are often used in the initial phase,
because of the lack of data on demand history. In the normal phase, causal methods
also have an important role, if data on explanatory variables are available. Some
models have been proposed, for example, that link the sales rate with the renewal
function associated with the part replacement in order to derive the demand for
spares (Blischke and Murthy, 1994). If such data are not available, then time-series
methods are used, usually based on exponential smoothing. In the final phase,
regression-based extrapolations have been recommended, assuming an exponential
decline of demand.
A further aim of classification, in the normal phase, is to determine the most
appropriate forecasting method. By examining the sources of intermittence and
erraticness of demand, it may be possible to identify parts with few customers and
to forecast using advance information or to predict responses to promotional
activity. For SKUs where this is not possible, two approaches have been proposed:
bootstrapping and distribution-based. In the former case, no distributional
 Forecasting and Inventory Management for Service Parts 25

assumptions are made and the lead-time distribution of demand is generated by re-
sampling from previous observations. In the latter case, a demand distribution must
be determined and its parameters estimated. Classification by the shape of the
demand distribution allows the system to determine whether the Poisson, the
compound Poisson or some other distribution should be used. Classification by
demand frequency and by demand size variability allows the system to choose
between smoothing methods such as Single Exponential Smoothing (for non-
intermittent, non-lumpy data) and methods such as Croston’s (for intermittent or
lumpy data).
The time-series method that should be employed for service parts depends on
the characteristics of the demand pattern. For non-intermittent demand, exponential
smoothing methods should be employed, with appropriate variants being chosen
for trended, damped trended and seasonal data. For intermittent demand series,
Croston’s method is the standard approach and has been adopted by a number of
forecast packages. The performance of Croston’s method can be improved by
applying an appropriate adjustment factor to reduce the bias of the forecast. This
has been shown by Eaves and Kingsman (2004) and by Syntetos and Boylan
(2006) to improve the inventory performance of the system.
The performance of forecasting methods for non-intermittent demand can be
assessed directly using measures such as the Mean Absolute Percentage Error
(MAPE) and the Mean Absolute Scaled Error (MASE). However, the MAPE is not
defined for intermittent series. For an individual series, forecast accuracy may be
assessed using the Mean Error and the Mean Absolute Error: Mean Demand ratio.
The latter measure has the shortcoming of producing unreliable results for trended
data, but trend is often barely perceptible in intermittent series. Alternatively, a
Geometric Mean (across series) of the Arithmetic Mean (across time) of the
Absolute Errors (GMAMAE) can be used. Statistical measures of forecast accuracy
should not be used alone, however, since optimization of forecast accuracy does
not necessarily lead to optimization of inventory performance. Inventory measures
assessing inventory costs and service level, particularly the fill rate, should also be
considered. Taken together, forecast accuracy and inventory measures provide the
manager with a comprehensive overview of the system’s performance.
Empirical evidence on the forecasting and inventory management of service
parts is not extensive, but has grown in recent years. There is good empirical
support for both compound Bernoulli and compound Poisson demand distributions.
The Negative Binomial Distribution has been found to be a good fit to many parts’
distributions, although some SKUs do not appear to be well represented by any
standard statistical distribution. Simple forecasting methods often work well, with
the Simple Moving Average being a good benchmark method for service parts with
intermittent demand. Croston’s method and its bias-reduced variants, including the
Syntetos-Boylan Approximation (SBA), should be considered. The SBA method
has been shown to perform well from both forecasting and inventory management
perspectives.
In summary, there has been substantial progress in research on forecasting for
inventory management of service parts over recent decades. Three challenges
remain: for researchers to resolve theoretical inconsistencies and develop more
powerful methods, for software manufacturers to reflect the state of the art in their
26 J.E. Boylan and A.A. Syntetos

packages, and for both researchers and software developers to work with
practitioners to broaden the base of empirical evidence in this field.

20.8 References
Bartezzaghi E, Verganti R, Zotteri G, (1996) A framework for managing uncertain lumpy
demand. Paper presented at the 9th International Symposium on Inventories, Budapest,
Hungary
Blischke WR, Murthy DNP, (1994) Warranty cost analysis. Marcel Dekker, Inc., New York
Boylan JE, (1997) The centralisation of inventory and the modelling of demand.
Unpublished Ph.D. Thesis, University of Warwick, UK
Boylan JE, Johnston FR, (1994) Relationships between service level measures for inventory
systems. Journal of the Operational Research Society 45: 838-844
Boylan JE, Syntetos AA, (2003) Intermittent demand forecasting: size-interval methods
based on average and smoothing. Proceedings of the International Conference on
Quantitative Methods in Industry and Commerce, Athens, Greece
Boylan JE, Syntetos AA, (2006) Accuracy and accuracy-implication metrics for intermittent
demand. Foresight: the International Journal of Applied Forecasting 4: 39-42.
Boylan JE, Syntetos AA, Karakostas GC, (2006) Classification for forecasting and stock-
control: a case-study. Journal of the Operational Research Society: in press
Brown RG, (1963) Smoothing, forecasting and prediction of discrete time series. Prentice-
Hall, Inc., Englewood Cliffs, N.J.
Brown RG, (1967) Decision rules for inventory management. Holt, Reinhart and Winston,
Chicago
Burgin TA, (1975) The gamma distribution and inventory control. Operational Research
Quarterly 26: 507-525
Burgin TA, Wild AR, (1967) Stock control experience and usable theory. Operational
Research Quarterly 18: 35-52
Croston JD, (1972) Forecasting and stock control for intermittent demands. Operational
Research Quarterly 23, 289-304
Croston JD (1974) Stock levels for slow-moving items. Operational Research Quarterly 25:
123-130
Department of Defense USA, (1980) Procedures for performing a Failure Mode, Effects and
Criticality Analysis. MIL-STD-1629A
Dunsmuir WTM, Snyder RD, (1989) Control of inventories with intermittent demand.
European Journal of Operational Research 40: 16-21
Eaves AHC, (2002) Forecasting for the ordering and stock holding of consumable spare
parts. Unpublished Ph.D. thesis, Lancaster University, UK
Eaves A, Kingsman BG, (2004) Forecasting for ordering and stock holding of spare parts.
Journal of the Operational Research Society 55: 431-437
Ehrhardt R, Mosier C, (1984) A revision of the power approximation for computing (s, S)
inventory policies. Management Science 30: 618-622
Fildes R, (1992) The evaluation of extrapolative forecasting methods. International Journal
of Forecasting 8: 81-98
Fortuin L, (1980) The all-time requirements of spare parts for service after sales –
theoretical analysis and practical results. International Journal of Operations and
Production Management 1: 59-69
Fortuin L, Martin H, (1999) Control of service parts. International Journal of Operations and
Production Management 19: 950-971
 Forecasting and Inventory Management for Service Parts 27

Friend JK, (1960) Stock control with random opportunities for replenishment. Operational
Research Quarterly 11: 130-136
Gallagher DJ, (1969) Two periodic review inventory models with backorders and stuttering
Poisson demands. AIIE Transactions 1: 164-171
Gardner ES, (1990) Evaluating forecast performance in an inventory control system.
Management Science 36: 490-499
Gardner ES, Koehler AB, (2005) Correspondence: Comments on a patented bootstrapping
method for forecasting intermittent demand. International Journal of Forecasting 21:
617-618
Ghobbar AA, Friend CH, (2002) Sources of intermittent demand for aircraft spare parts
within airline operations. Journal of Air Transport Management 8: 221-231
Ghobbar AA, Friend CH, (2003) Evaluation of forecasting methods for intermittent parts
demand in the field of aviation: a predictive model. Computers and Operations Research
30: 2097-2014.
Hoover J, (2006) Measuring forecast accuracy: omissions in today’s forecasting engines and
demand-planning software. Foresight: the International Journal of Applied Forecasting
4: 32-35
Hua ZS, Zhang B, Yang J, Tan DS, (2006) A new approach of forecasting intermittent
demand for spare parts inventories in the process industries. Journal of the Operational
Research Society: in press.
Hyndman RJ, (2006) Another look at forecast-accuracy metrics for intermittent demand.
Foresight: the International Journal of Applied Forecasting 4: 43-46
Janssen FBSLP, (1998) Inventory management systems; control and information issues.
Published Ph.D. thesis, Centre for Economic Research, Tilburg University, The
Netherlands
Johnston FR, (1980) An interactive stock control system with a strategic management role.
Journal of the Operational Research Society 31: 1069-1084
Johnston FR, Boylan JE, (1996) Forecasting for items with intermittent demand. Journal of
the Operational Research Society 47: 113-121
Johnston FR, Harrison PJ, (1986) The variance of lead-time demand. Journal of the
Operational Research Society 37: 303-308
Kalchschmidt M, Verganti R, Zotteri G, (2006) Forecasting demand from heterogeneous
customers. International Journal of Operations and Production Management 26: 619-638
Kwan HW, (1991) On the demand distributions of slow moving items. Unpublished Ph.D.
thesis, Lancaster University, UK
Makridakis S, (1993) Accuracy measures: theoretical and practical concerns. International
Journal of Forecasting 9: 527-529
Makridakis S, Hibon M, (2000) The M3-Competition: results, conclusions and implications.
International Journal of Forecasting 16: 451-476
Naddor E, (1975) Optimal and heuristic decisions in single and multi-item inventory
systems. Management Science 21: 1234-1249
Quenouille MH, (1949) A relation between the logarithmic, Poisson and negative binomial
series. Biometrics 5: 162-164
Rao AV, (1973) A comment on: Forecasting and stock control for intermittent demands.
Operational Research Quarterly 24: 639-640
Ritchie E, Kingsman BG, (1985) Setting stock levels for wholesaling: performance
measures and conflict of objectives between supplier and stockist. European Journal of
Operational Research 20: 17-24
Ronen D, (1982) Measures of product availability. Journal of Business Logistics 3: 45-58
Sani B, Kingsman BG, (1997) Selecting the best periodic inventory control and demand
forecasting methods for low demand items. Journal of the Operational Research Society
48: 700-713
 28 J.E. Boylan and A.A. Syntetos

Shale EA, Boylan JE, Johnston FR, (2006) Forecasting for intermittent demand: the
estimation of an unbiased average. Journal of the Operational Research Society 57: 588-
592
Shenstone L, Hyndman RJ, (2005) Stochastic models underlying Croston’s method for
intermittent demand forecasting. Journal of Forecasting 24: 389-402
Silver EA (1970) Some ideas related to the inventory control of items having erratic demand
patterns. CORS Journal 8: 87-100.
Silver EA, Pyke DF, Peterson R, (1998) Inventory management and production planning
and scheduling (3rd edition). John Wiley & Sons, New York
Snyder R, (2002) Forecasting sales of slow and fast moving inventories. European Journal
of Operational Research 140: 684-699
Strijbosch LWG, Heuts RMJ, van der Schoot EHM, (2000) A combined forecast-inventory
control procedure for spare parts. Journal of the Operational Research Society 51: 1184-
1192
Syntetos AA, (2001) Forecasting of intermittent demand. Unpublished PhD Thesis,
Buckinghamshire Chilterns University College, Brunel University, UK
Syntetos AA, Boylan JE, (2001) On the bias of intermittent demand estimates. International
Journal of Production Economics 71: 457-466
Syntetos AA, Boylan JE, (2005) The accuracy of intermittent demand estimates.
International Journal of Forecasting 21: 303-314
Syntetos AA, Boylan JE, Croston JD, (2005) On the categorization of demand patterns.
Journal of the Operational Research Society 56: 495-503
Syntetos AA, Boylan JE (2006) On the stock control performance of intermittent demand
estimators. International Journal of Production Economics 103: 36-47
Teunter RH, (1998) Inventory control of service parts in the final phase. Published PhD
Thesis, University of Groningen, The Netherlands
Teunter RH, Fortuin L, (1998) End-of-life-service: a case-study. European Journal of
Operational Research 107: 19-34
Vereecke A, Verstraeten P, (1994) An inventory management model for an inventory
consisting of lumpy items, slow movers and fast movers. International Journal of
Production Economics 35: 379-389
Ward JB, (1978) Determining re-order points when demand is lumpy. Management Science
24: 623-632
Watson RB, (1987) The effects of demand-forecast fluctuations on customer service and
inventory cost when demand is lumpy. Journal of the Operational Research Society 38:
75-82
Willemain TR, Smart CN, Shockor JH, DeSautels PA, (1994) Forecasting intermittent
demand in manufacturing: a comparative evaluation of Croston’s method. International
Journal of Forecasting 10: 529-538
Willemain TR, Smart CN, Schwarz HF, (2004) A new approach to forecasting intermittent
demand for service parts inventories. International Journal of Forecasting 20: 375-387
Williams TM, (1984) Stock control with sporadic and slow-moving demand. Journal of the
Operational Research Society 35: 939-948

View publication stats