You are on page 1of 7

Integration approaches of forecasting methods

selection with inventory management indicators in


the case of spare parts supply chain
LAZRAK Adnane*1

CASTANIER Bruno2, LEMOINE David3

Ecole des Mines de Nantes/GE Healthcare


*1
adnane.lazrak@ge.com

Ecole des Mines de Nantes


bruno.castanier@mines-nantes.fr,
3
david.lemoine@mines-nantes.fr
2

HEIDSIECK Robert4, THENOT Cyrille5


GE Healthcare
4

robert.heidsieck@med.ge.com,
cyrille.thenot@ge.com

Abstract The main scope of this paper is to improve


management policies for spare parts, within the context of
centralized management and more particularly with regard to
forecast methods. The specificity of low and erratic demand does
not allow the use of conventional approaches of forecasting.
Moreover, the associated performance measurements, based on
purely statistical indicators, are not adapted to the context of
supply chain management. Here we propose two new
performance analysis approaches seeking to combine the
statistical performance of forecasting methods and inventory
management performance. A comparative analysis of forecasting
methods on real data enable the definition of a strategy for
selecting the appropriate method when integrated with spare
parts inventory management model using a continuous review
(s, S).
Keywords Forecast methods; inventory management, (s, S)
policy; spare parts supply chain; supply chain strategy;
performance analysis.

I.

INTRODUCTION

Supply chain management is a key element of industrial


management, enabling to improve the time, cost and quality
triangle for the company.
Although it seems just one particular aspect of supply chain
management, spare parts supply chain has been subject to
particular interest and independent literature [1].
The demand forecast in the case of the spare parts supply
chain, presents special difficulties [2]. Indeed, analyses of
demand patterns show that they are usually low and sporadic.
Moreover, the spare parts criticality to insure the best
availability of systems to the end user may cause some
variability in this demand.
The classification of these demand forecasting methods
rests on purely statistical considerations without taking into
account service level and inventory level goals after their
integration with the process of inventory management. This

can later result in a bad design of inventory management policy


including the calculation of replenishment thresholds.
We propose in this paper, on the basis of a real industrial
case that we briefly describe below, to focus on the
construction and the analysis of new approaches for forecasting
methods selection based on a set of performance criteria that is
inventory management oriented, while integrating the main
characteristics of the part category.
GE Healthcare is the medical branch of conglomerate
General Electric, one of the global leaders in sales and services
of medical systems notably those of medical imaging.
On account of the criticality of its products (medical
devices) and the technological characteristics of its
components, GE Healthcare offers a maintenance service to its
customers. The main objective of this service is to ensure the
reliability of its products (reducing the failure rate occurrence),
while reducing at the same time periods of unavailability.
GE Healthcare needs to clearly control the complete spare
parts supply chain and distribution processes, because of the
criticality and cost of its parts and the geographical dispersion
of its customers.
GPRS Global Parts & Repair Solutions department
within the company is responsible for service parts. This covers
the whole of the parts supply chain management, warehouses
management, choice of transport, inventory management
policies throughout the network, and repair strategies for
defective parts.
In this overall context GPRS is constantly looking for
efficiency and improvement to these processes in order to
reduce costs and ensure customer satisfaction.
One of the improvement levers identified is the quality of
demand estimation for spare parts. The analysis of the practices
of the planning team based on the use of commercial software
has led to the following observations:

- The input data for the models used in demand forecasting


contained in the software are mainly based on customer orders.
This then uses the classical models of demand forecasting.
However, these models are known to be ineffective for low or
erratic demand, the latter properties are generally checked for
spare parts. Finally, the operational environment of GPRS does
not currently allow migration to new reliability oriented
approaches [3].
- It is not possible today to qualify a demand forecasting
model based on parts inventory levels or in terms of demand
satisfaction. This is partly explained by the complexity of the
spare parts supply/distribution process.
The second observation motivates us to propose new
measurements for the performance of forecast methods adapted
to the context of spare parts inventory management. These
measurements will of course assess the error between forecast
and demand but also the impact on the final indicators in
fulfillment and in inventory. We keep as final indicators the
combination between the quantity of "backorders" (stock out)
and the inventory level. Performance indicators can also align
other aspects of supply chain requirements such as lead time,
the specificity of the part category, the annual
inventory/fulfillment strategies and goals.
The problem of the selection of demand forecasting models
on such criteria has been little addressed in the literature.
Performance analysis restricted to statistical considerations is
not sufficient and the definition of specific criteria remains
critical.
In this paper, we propose to define a selection process for
forecasting models defined in two fields: statistical
performance and inventory management performance. We seek
to propose new approaches for the evaluation of the
performance of these criteria and a comparative analysis within
the context of spare parts inventory management using a
classical continuous review policy (s, S) under backorders.
The organization of the paper is as follows. First, a
literature review is presented in Section 2 on the themes of both
statistical accuracy measurements for the selection of spare
parts forecast models and the integration of these models with
inventory management policies. The purpose of Section 3 is the
construction of two new approaches for forecast methods
selection based on inventory management criteria. The
comparative analysis of these processes on real data is
proposed in Section 4. The conclusion, Section 5, will allow a
discussion of our proposals and define areas for future works.

II.

LITERATURE REVIEW

In the context of spare parts management, forecasting


methods and accuracy measurements has been made the object
of specific literature because of the characteristics of this area.
Spare parts demand does not depend on customer behavior
(such as that frequently found in a sales context), but is more
due to a maintenance policy allied to failure issues and other
risks inherent in the systems manufacturing [1]. The needs
generated are therefore much more erratic and contain a high
variability [2]. Hence, specific methods of forecasting have
been developed [4], [5], [6].
Ad-hoc statistical indicators have also been proposed to
measure the performance of these methods [7], [8]. Reference
[9] highlighted that the choice of the measurement strongly
impacts the selection of the forecast method. Reference [10]
Listed different types of statistical forecast measurements.
Table I names some of standard measurements found in
related literature:
-

is the forecast for period i.

is the real demand of period i.

TABLE I.

is the number of comparison periods between


forecast and real demand.
CLASSICAL FORECAST ACCURACY MEASUREMENTS
Measurement Name

Formula

Mean Squared Error (MSE)

Mean Absolute Deviation (MAD)

Mean Absolute Percent Error (MAPE)

100

However, all these indicators are based on statistical


analysis of the difference between the real demand and
forecasts generated by the forecast method used. Thus, a
forecast method may be preferred over others based on one or
more of these measurements and forecasts generated by this
method can be used as input data for the inventory
management process as shown in Figure 1.

Part classification

Forecast
Process

Forecast methods

Evaluation of the forecast methods according


to statistical indicators

Selection of the best forecast method

Inventory
management
Process

Inventory management model

Performance analysis according to inventory


criteria

Fig. 1.

Classical approach of forecast methods selection process

It is important to note that, in this configuration (figure 1),


the forecasts generated are not being challenged at the level of
inventory management process (these are only the input data)
and also serve in the setting of the inventory management
model (For example in the calculation of safety stock). Thus,
besides the importance of the choice of performance
measurements for the selection of the appropriate forecast
method [9], we need to ask the question regarding the lack of
consideration of the inventory management performance
process in this selection.
Then, we propose to carry out the selection after integration
with the inventory management model.
Inventory management models literature generally follows
one classification into two groups depending on the type of
review: a periodic review a so-called (T, r, Q) policy [11] or a
continuous review a so-called (s, S) policy [12].
It also differentiates between two orders management
policies: "lost sales", or "Backorders".
Due to contractual aspect of the after-sales service,
inventory management models for spare parts are based on
"Backorders" model type rather than "lost sales" model [13].
Moreover, thanks to their advantage in terms of
responsiveness, continuous review models (s, S) permit a better
response to the erratic pattern of demand [14].
Inventory management models based on forecasting, used
in this paper, are different from those based on demand [15]:
They use, forecast error in safety stock calculation. Reference
[16] tested the effect of forecast errors on safety stock size.

References [17], [18] integrated forecast methods with


inventory management models and evaluated this combination
with cost and service level performance indicators.
However, these integration papers do not propose new
approaches of forecast methods selection using both inventory
management performance and according to the part category.
They do not propose new measurements that combine those
performance criteria.
Therefore, in the following section we propose to fill in this
missing link.
III.

PROPOSED APPROACHES

In this section, we detail the two approaches with which we


propose to address the problem of not taking into account
aspects of inventory management in the choice of forecast
methods.
The first approach consists of keeping the statistical
performance indicators for forecasting while starting to
consider inventory management aspects, the second one lies in
the total integration of the selection process for forecasting and
inventory management.
A. Inventory management oriented forecast accuracy
measurements (Approach 2)
In this sub-section, we start from the classical scheme
shown in Figure 1 but we define new selection measurements
of forecast methods containing inventory management oriented
components: Therefore the sub-process "Evaluation of forecast
methods according to statistical indicators" in figure 1 process
will involve these new measures.
1.

An inventory management Risk-based measurement

In the statistical indicators we have shared previously


(Table I), the accuracy of the forecasts is measured either in
mean squared error or mean absolute error. However, in
inventory management context, it is of course important to
know if the forecast is accurate, but also whether it exceeds the
demand (and in this case, inventory excess risk) or if it is
underestimating demand (in this case, backorders generating
risks). Hence, we propose a new measurement that will be the
convex combination of two indicators one to reflect the
forecasts overestimation of the real demand and the other its
underestimation.
For the present, we take the notations used in Section 2.
As an indicator that reflects the risk of the inventory excess,
we define the Mean Positive Error (MPE):
0,
As an indicator that reflects the risk of the generation of
backorders, we define the Mean Negative Error (MNE):
0,

Thus we define a measure we call Inventory Orientation


Error (IOE) as the convex combination of these two
components:
.

According to value, IOE will translate the necessary


guidelines for the decision makers, or according to the
classification of the part.
-

If 0,5, the emphasis is on improving service level


(The goal is to reduce backorders).

If 0,5, the emphasis is on reducing the inventory


excess risk.

If

2.

A lead time-based forecast accuracy measurement

1
1

Part classification

Forecast methods

Inventory management model

Performance analysis according to


weighted inventory criteria

0,5, we find back MSE.

In the forecast accuracy measurements that we find in the


literature as well as those we have presented previously, the
forecasts and current demands are compared period by period.
However, in inventory management, purchased orders are often
done in batches to minimize the related logistics costs.
Generally, the size of this lot covers demand that may occur
during the lead time. It is therefore important to measure the
accuracy of the aggregated forecast on the lead time compared
to real demand during lead time, to ensure that the quantities
ordered, adequately cover the demand. Then, we propose a new
measure called Lead Time Forecast Error (LFTE) which we
calculate as follows:

Where :
-

Integrated
process of
selection of
the forecast
methods
according to
inventory
management

is the forecast calculated at period i for period t.


is the lead time value.
is the number of comparison periods between
forecast and real demand.
is the real demand for period i.

B. Integrated Process for the Selection of forecast method


with inventory management indicators (Approach 3)
The underlying idea to this integration is the direct use of
indicators of inventory management to select the best
forecasting method. This integration is given in Figure 2.

Selection of the best forecast


method

Fig. 2.

Integrated Assessment Approach

In this approach, we propose to test all the forecasting


methods and use a historical demand-based simulation of the
impact in terms of inventory management performance based
on forecasts by comparing them to the real demand occurred
during this period. This impact will be directly measured
through indicators used by inventory management, and then the
forecasting method having the best performances will be
selected.
To be able to compare this approach to those presented
above, we use a measurement constructed from the most
traditional performance indicators of inventory management
that will be a reinterpretation of the IOE measure in this
context.
Thus we denote:
-

: The number of forecast methods tested.


: is the inventory level (On Hand) at period i
using the forecast method j.
: is the backorder level at period i using the
forecast method j.

We define the Measurement of the Relative Inventory to


the Forecast method j in the following way:

And Measurement of the Relative Backorders quantity to


the forecast method j:

These two components allow the respective measurement


of the levels of inventory and backorder generated by the use of
the forecast method during the considered horizon.

We then define our indicator called Relative Inventory


Performance Metric, for forecast method j as a convex
combination of these two metrics:
.

- Two conventional methods, rather adapted to frequent


demand:
- The Moving Average method (MA): the forecast of this
method is an average of the demand over a specific period.

Like the IOE indicator, the value of will translate the


necessary guidelines for the decision makers, or according to
the classification of the part.

- The Simple Exponential Smoothing Method (SES): The


forecast of this method is a convex combination of the last
forecast and the last demand.

C. Comparison between the three approaches of forecast


methods selection
Below, Table II summarizes the proposed approaches and
compares them with classical approach.

- Two methods adapted to low and sporadic demand,


particularly verified in the context of spare parts [2]:

TABLE II.

Advantage

Disadvantage

Usage
context

APPROACHES COMPARISON TABLE


Approach 1
(classical)
Selecting
the
forecast
method
with
the
best
statistical
performance

No consideration of
inventory
management
criteria
in
the
selection

Need
only the
method with the
best
statistical
accuracy

Approach 2
(proposed)
Consideration of
inventory
management
performance risk
in the selection.
Without the need
to integrate with
an
inventory
management
model
Selection
by
risks only and
not by direct
measure of these
criteria

Approach 3
(proposed)
Selection of the
forecast method
with the criteria
of
inventory
management

Need the method


with the best
performance in
inventory
management, but
inability to use
the approach 3
for
IT
constraints

Need
the
method
with
the
best
inventory
management
performance

- The Croston Method (CR) [4]: The forecast of this


method is based on a double smoothing: the first is the demand
size; the second is the interval between demands.
- The Synthetos and Boylan Method (SBA) [5]: is a variant
of the Croston method that eliminates its statistical bias.
The model of inventory management used is an (s, S)
policy under backorders, with regard to forecasts and forecast
error. We recall that s defines the safety stock and S the
replenishment threshold.
We assume a constant lead time (LT).

No
explicit
selection
by
statistical
criteria. But the
statistical error
is contained in
the settings of
safety stock

IV. RESULTS
A. Experimental Design
The objective of this sub-section is the characterization of
the Design of experiments (DOE) in order to illustrate the
performance of measures of selection. This design is defined
by a set of forecast methods, a set of policies for managing
stock and business priorities in terms of service level or overstocking.
For brevity, the analysis will be reduced to the single (s, S)
policy of inventory management, the most common policy in
the context of spare parts.
For forecasting methods, we choose to consider two types
of methods characterizing demand type. For each type, we will
use two methods. We will test:

We will consider a fixed safety stock level and a dynamic


replenishment level updated each period based on new
generated forecasts, denoted (s, Si) , with :

level.

is the -score, with 1

defining the required service

Backorders management is as follows: when an order


arrives from the supplier, the backorders remaining from
previous periods are closed first before satisfying the demand
for the current period.
In terms of business priority, if we will focus on priority on
inventory more than service level then we will take
0.7 ,
if more on service level than inventory we will take
0.3.
We test this DOE on an extract from a database from
GPRS. This database is composed of 36 monthly demands of
500 spare parts references. The first 12 months will serve for
models calibration. The remaining 24 months will serve for the
analysis and comparison of forecast methods by different
measures.
Finally, we assume a constant lead time of 3 months and a
required service level of 95%.
B. Results
We recall here that approaches 1 and 2 are unable to update
the inventory management policy contrary to the measures of
the third approach.
In the following table III, we represent, for each column,
the percentage of 500 references of the database for which the
forecast method was better evaluated using the measure of the
column.

TABLE III.

RESULTS IN PERCENTAGE OF SELECTION OF FORECAST

3.2. The case where the history is not sufficient:

METHODS

Approach 1
(Section II)
MSE
MA
SES
CR
SBA
Total

6.4
3.8
26.6
63.2
100%

Approach 2
(Section III.A)
LTFE
.
.
14.4
1.2
14
70.4
100%

27
5
43
25
100%

33
9.6
18.8
38.6
100%

Approach 3
(Section III.B)
.

23
8
27
42
100%

27
10
29
34
100%

The first column of data clearly illustrates the advantage in


terms of statistical performance of specific forecast methods
for low and erratic demand, principally the SBA method.
However, the following results, and especially the
significant improvement in results of traditional approaches of
forecasting, show that there is not a unique forecast method
that outperforms the others in all cases.
We are able, for example, to note that the dominance of the
SBA (column 1) decreases significantly when the supply chain
concerns are reinforced (approach 3). This conclusion
reinforces the interest of our study because it does not confirm
the intuitive idea that the search for the best adequacy for
forecasting to real demand necessarily leads to improve
performance in terms of inventory management. However, It
remains true that in more than two thirds of cases (sum of CR
and SBA lines), this research on statistical performance
adjustment translates into a benefit at the level of the supply
chain.
One is able to note more specific results on each of the
proposed approaches. One will note also that the improvement
provided by the SBA Method to the CR Method is significant
in the case of cost orientation but less relevant in the case of
service orientation.
In view of these results, we propose to follow this general
process for the selection of forecast method:
1. Prioritize the supply chain in terms of expectations and
the class of the part (cost, criticality ...). This sets the selected
approach and the associated criterion (choice of measurement
and associated ).
2 . Characterize the reference studied in terms of demand
category and the maturity level. This maturity is determined
depending on the size of the available historical data and the
life cycle of the reference.
3 Select the method for demand forecasting:

a. Identify the category that is "closest". This identification


can be done mainly on the basis of commercial and
technological data in order to study similarities in terms of
technology, functioning conditions and especially demand
category (smooth , erratic, ... ).
b. Apply the most efficient method of forecasting in the
identified category.
c. Update the level of maturity of the reference.
V.

In this paper, we analyzed the performance of several


demand forecast methods on real data sets of spare parts based
on several performance criteria. These new criteria allow better
consideration of the business priorities such as the performance
of the supply chain process. The most important result of this
analysis is that there is not a single forecast method
outperforming the others in all cases. It is not possible to
choose one in advance.
Therefore, we have proposed a general process for the
selection of the best forecast method. It is based on a
systematic evaluation of the performance using several
measures/weights for each reference. In addition, it offers a
better match between forecast method and business priorities in
terms of fulfillment and inventory management defined for
each reference. This search for specificity in the choice of
forecasting method in relation to the characteristics of the
reference seems to us quite within the reach of industrial
challenges through the customizing of their products and
services.
The further works remain numerous. The first that we are
able to cite is related to the characterization of the concept of
parts maturity that we have just introduced by linking it to the
lifecycle of the demand for each reference. This
characterization will certainly go through a quantification of
parameters such as, for example, the size of the historicdemand available for reference, but also an extension of its
definition by taking into account the economic lifecycle of the
same reference.
A second perspective is the enrichment of the forecast
methods database as well as parts inventory management
policies. This last point may lead us to define also new specific
measures related to Approach 3.
REFERENCES
[1]

3.1. The case where the history is sufficient:


a. Test all of the forecasting methods based on the criteria
set out in point 1. This analysis can be easily automated and
requires no effort by the analyst (Supply leader). It may in
some cases even be transparent.
b. Choose the most efficient method of forecast and update
the evaluations of each method in the database "References
categories".

CONCLUSION

[2]

[3]

[4]

W.J. Kennedy, J .Wayne Patterson, and L.D. Fredendall, An overview


of recent literature on spare parts inventories, International Journal of
Production Economics, 76 (2), pp. 201215, 2002.
A.A. Syntetos, J.E. Boylan and J.D. Croston, On the categorization of
demand patterns, Journal of the Operational Research Society, 56 (5),
pp .495503, 2005.
A.B.M. Zohrul Kabir,and A.S. Al-Olayan, A stocking policy for spare
part provisioning under age based preventive replacement, European
Journal of Operational Research, 90, pp. 171181, 1996.
J.D. Croston, Forecasting and stock control for intermittent demands,
Journal of the Operational Research Society, 23 (3):289303, 1972.

[5]

A.A. Syntetos, and J.E. Boylan, On the Bias of Intermittent Demand


Estimates, International Journal of Production Economics, 71 (1-3),
pp. 457-466, 2001.
[6] W. Romeijnders, R.H. Teunter, and W. Van Jaarsveld, A two-step
method for forecasting spare parts demand using information on
component repairs, European Journal of Operational Research, 220 (2),
pp. 386393, 2012.
[7] J.S. Armstrong, and F. Collopy, Error measures for generalizing about
forecasting methods: Empirical comparisons, International Journal of
Forecasting, 8(1), pp. 69-80, 1992.
[8] A.A. Syntetos, and J.E Boylan, The Accuracy of Intermittent Demand
Estimates, International Journal of Forecasting, 21, pp. 303 314,
2005.
[9] V. Varghese, and M.Rossetti, A classification approach for selecting
forecasting techniques for intermittent demand, In proceedings of the
Industrial engineering research conference, Vancouver Canada, IEEE,
2008.
[10] R.J. Hyndman, and A.B. Koehler, Another look at measures of forecast
accuracy, International Journal of Forecasting, 22, pp. 679688, 2006.
[11] A. J. Clark, and H. Scarf, Optimal policies for a multi echelon
inventory problem, Management science, 6(4), pp. 475-490, 1960.

[12] A. F. Veinott, Jr, On the Opimality of (s,S) Inventory Policies: New


Conditions and a New Proof, SIAM Journal on Applied
Mathematics, 14(5), 1067-1083, 1966.
[13] D. Louit, R Pascual, D. Banjevic, and A. K. Jardine, Optimization
models for critical spare parts inventoriesa reliability approach,
Journal of the Operational Research Society, 62(6), pp. 992-1004, 2010.
[14] S. Axster, Continuous review policies for multi-level inventory
systems with stochastic demand, Handbooks in operations research and
management science, 4,pp. 175-197, 1993.
[15] S. C. Graves, A single-item inventory model for a nonstationary
demand
process,
Manufacturing
&
Service
Operations
Management, 1(1), pp. 50-61, 1999.
[16] E. A. Silver, A modified formula for calculating customer service
under continuous inventory review, AIIE Transactions, 2(3), 241-245,
1970.
[17] E. Levn,and A. Segerstedt, Inventory control with a modified Croston
procedure and Erlang distribution, International Journal of Production
Economics, 90(3), 361-36, 2004.
[18] L.Tiacci, and S. Saetta, An approach to evaluate the impact of
interaction between demand forecasting method and stock control policy
on the inventory system performances, International Journal of
Production Economics, 118(1), 63-71, 2009.