On the Categorization of Demand Patterns Author(s): A. A. Syntetos, J. E. Boylan and J. D.

Croston Source: The Journal of the Operational Research Society, Vol. 56, No. 5 (May, 2005), pp. 495503 Published by: Palgrave Macmillan Journals on behalf of the Operational Research Society Stable URL: http://www.jstor.org/stable/4102103 . Accessed: 07/08/2013 14:51
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Palgrave Macmillan Journals and Operational Research Society are collaborating with JSTOR to digitize, preserve and extend access to The Journal of the Operational Research Society.

http://www.jstor.org

This content downloaded from 146.155.157.12 on Wed, 7 Aug 2013 14:51:14 PM All use subject to JSTOR Terms and Conditions

Journal of the Operational Research Society (2005) 56, 495-503

2005 OperationalResearchSociety Ltd.All rightsreserved. 0160-5682/05 $30.00 www.palgrave-journals.com/jors

On the categorization of demand patterns

AA Syntetos'*, JE Boylan2 and JD Croston3 Chilterns Manchester, College,Bucks,UK;and 'University of Salford,Greater University UK;2Buckinghamshire Herts, UK Systemsand StatisticalConsultant, 3lndependent Thecategorization of alternative demand facilitates of a forecasting theselection method andit is an essential patterns element of manyinventory in the inventory control software Thecommon controlsoftware industry packages. practice is to arbitrarily thosedemand andthenproceed to selectan estimation andoptimize the procedure categorize patterns forecastparameters. methodscan be directlycompared, based on some theoretically Alternatively, forecasting error for the purpose andthendefine of establishing of superior the demand measure, quantified regions performance basedon the results. It is thisapproach in thispaperandits application thatis discussed is demonstrated patterns by Croston's method andan alternative to Croston's estimator of EWMA, considering by thefirsttwoauthors developed this paper.Comparison results arebasedon a theoretical of themeansquare errordueto its mathematically analysis tractable nature. Thecategorization rulesproposed in terms areexpressed of theaverage interval inter-demand andthe coefficient of variation of demand sizes.Thevalidity of the results is testedon 3000real-intermittent demand squared dataseriescomingfromthe automotive industry.

Journalof the Operational ResearchSociety (2005) 56, 495-503. doi:10.1057/palgrave.jors.2601841

Published online25 August2004

intermittent demand Keywords: control; categorization; forecasting; inventory

Introduction The categorization of alternative demand patterns is an essential element of many inventory control software packages. The common practicein the softwareindustryis to arbitrarily categorizedemandpatternsand then select an estimationprocedureand stock control method in order to forecast future requirementsand manage stock efficiently. For example,certainarbitrary cutoff values may be given to the numberof demand occurringperiods in a year, average demand per unit time period and standarddeviation of the demand sizes in order to define demand patterns as slow, intermittent,lumpy, fast, etc. In this paper, we are concernedwith the selectionof the most appropriate estimation procedure as a first step towards a more systematic and meaningful approach to the categorizationproblem. Typically,for fast moving items, analysis of the demand data series will lead to the selection of an appropriate forecastingmodel (or a class of models) and optimizationof the forecast parameterswill be based on minimizing the mean squareerror(MSE), Bayesianrecursionor some form of residualauto-correlationanalysis. For the remainderof the categories Moving Averages, simple Exponentially WeightedMoving Averages(EWMA) or Croston'smethod' are the most commonlychosenestimationprocedures. In the
AA Syntetos, The School of Management, University *Correspondence: of Salford, Maxwell Building, The Crescent,Salford, GreaterManchester M5 4 WT, UK.

E-mail: a.syntetos@salford.ac.uk

case of the sporadicand/or irregulardemand patterns, the sparsenessof data and/or the aggregatepolicies required, may result in an empirical determinationof the control parameters, so that, for example, a 13-period Moving Averageor a smoothingconstantequal to 0.15 are specified across a whole category. occurs at fixed time intervals(annually Re-categorization for example)and the sensitivityof the categorization rulesto outliers is usually examined so that the categorization schemedoes not allow productsto move from one category to the other simplybecausea few extremeobservationshave been recorded.Logical inconsistenciesare also explored so that demand for a stock keeping unit (SKU) is always classified in the intended category. However, the more consistent the categorization scheme is, the greater the numberof the demandcategoriesthat need to be considered. Theoreticaland practicalrequirements that should be taken into account when developing rules for the purpose of distinguishing between alternative demand patterns have been discussedin Williams.2 Ultimately, the objectiveof the categorizationexerciseis the selection of the most appropriate estimationprocedure. However, if this is the real objective it seems more meaningful to compare possible forecasting methods for the purpose of establishingregions of superiorperformance and then categorize the demand patterns based on the results.The theoreticalcomparisonsneed to be based on an accuracy measure, and the MSE is the most obvious candidate because of its mathematicallytractable nature.

This content downloaded from 146.155.157.12 on Wed, 7 Aug 2013 14:51:14 PM All use subject to JSTOR Terms and Conditions

496 Journal oftheOperational Research Vol. 5 56, No. Society

Direct comparisonof the theoreticalMSEs, over a fixed lead time, associated with any two estimation procedureswill result in establishingrules that indicate regions of superior of one method over the other. When there are performance more than two estimatorstaken into account there will be some overlapping areasregardingsuperiorperformance and the analysis will need to be extended in order to propose rules which are valid across all methods. We demonstrate our approach by considering three methods: (a) Croston's method, that has been specifically designed to deal with intermittence,(b) an alternativeto Croston's estimator,developed by the first two authors of this paper and (c) the more generally applicableEWMA. The categorization rules are expressedin termsof the series' inter-demand intervaland the squaredcoefficientof average variation of the demand sizes (when demand occurs). The validity of our results is tested on 3000 real-intermittent demanddata series.

Categorization schemes proposed in the literature Williams2proposed a method of categorizationof demand patterns based on an idea that is called variancepartition (we split the varianceof the demandduringlead time into its constituent parts). The purpose of categorizationwas the identification of the most appropriate forecasting and inventorycontrol methods for the resultingcategories. of the items takes place in accordancewith Categorization the matrix shown in Figure 1 (with the cutoff values being the result of managerialdecision). In the matrix shown below, 2 is the mean (Poisson) demand arrival rate, L the mean lead time duration and CV2(x)the squaredcoefficientof variationof demand sizes.

Lumpiness
CV2 (X)

1/,L indicatesthe numberof lead times betweensuccessive demands (how often demand occurs or how intermittent demand is). The higher the ratio, the more intermittent demand is. CV2(x)/)L indicates how lumpy demand is. Lumpiness depends on both the intermittence and the variability of the demand size, when demand occurs. The higher the ratio the more lumpy demand is: category Dlsporadic(lumpy);categoryD2-highly sporadic(lumpy).In that case we have few very irregulartransactions accompanied by highly variable demand sizes; category B-slow moving; others-smooth. Eaves3analyseddemand data from the Royal Air Force (RAF) and concluded that Williams' conceptual classification scheme did not adequately describe the observed demand structure. In particular, it was not considered sufficientto distinguisha smooth demand pattern from the remaindersimply on the basis of the transactionvariability. Consequently,a revisedclassificationscheme was proposed (see Figure 2) that categorizes demand based on the variabilityof the transactions'rate, demand size variability The line itemswith low transaction and lead time variability. into smooth and irregular sub-divided are variability size the to demand variability.The smooth and according demand categories are distinguishedfrom the slow-moving rest based on the varianceof the demand sizes, and the lead betweenerratic time varianceis used only for distinguishing and highly erraticdemand. The cutoff values were decided based on: (a) the characteristicsof the particulardemand data set and (b) sufficientsub-sample(within each demand class) size considerations.In particular, the cutoff points were as follows: transactionvariability-0.74; demand size variability--.10; lead time variability-0.53. CategoryA smooth; category C-irregular; category B-slow-moving; category Dl-erratic; category D2-highly erratic. The above-discussedcategorizationschemes are the only ones to appear in the academic literature that take into considerationalternativedemandpatternsand consequently distinguishbetween them for the purpose of selecting the most appropriate forecasting(and stock control) procedure. values have been arbitrarilychosen the cutoff Nevertheless, so that they make sense only for the particularempirical
sizevariability Demand

0.5 A 0.7 B LD2 2.8 D2 D1 0.74

Transaction B

C Intermittence

0.10
A C

Dl 0.53 Lead-time variability D2

variability

scheme. categorization Figure1 Williams'

scheme. Figure2 Eaves'categorization

This content downloaded from 146.155.157.12 on Wed, 7 Aug 2013 14:51:14 PM All use subject to JSTOR Terms and Conditions

AA et a/-Categorization ofdemand 497 patterns Syntetos

situationsthat were analysedin the corresponding pieces of research. Thereis no doubt that the criteriaused for categorization are meaningful.In fact, similarcriteriawill be used for the categorizationrules that will be proposed later on in this paper (with the differencethat lead times are assumedto be cutoff values assignedto constant).It is ratherthe arbitrary them that create certain doubts about the potential applicabilityof the proposedcategorizationschemesto any differentpossible context. An alternative approachto the demand categorization problem Johnston and Boylan4 compared Croston's intermittent demand estimationprocedurewith EWMA on theoretically generated demand data over a wide range of possible conditions. (Croston'sapproachto forecastingintermittent demand and the latest research findings regarding his method are presented in Appendix A). Many different average inter-demandintervals (negative exponential distribution), smoothing constant values, lead times and distributionsof the size of demand (negative exponential, were considered.The comparison Erlang and rectangular), exercise was extended to cover not only Poisson but also Erlang demand processes.The results were reportedin the form of the ratio of the MSE of one method over that of the other. For the different factor combinations used in this simulation experiment,Croston's method was superior to EWMA for inter-demand intervalsgreaterthan 1.25 review This result was first response to the question the periods. raised in Anders Segerstedt's paper:5 'When is best to separate the forecasts like Croston suggests and when is it best with traditional treatment (ie simple exponential smoothing)?'. Definition of intermittentdemand now results from a directcomparisonof possible estimationproceduresso that regions of relative performancecan be identified.It seems more logical indeed working in the following way: 1. Comparealternativeestimationprocedures. 2. Identifythe regionsof superiorperformance for each one of them. 3. Define the demand patterns based on the methods' ratherthan arbitrarily comparativeperformance defining demand patterns and then testing which estimation procedure performs best on each particular demand category. The approach discussed above is taken forward in the researchdescribed in this paper,in that the decisionruleswill be the outcome of a comparison between MSE performances.Thereare some significantdifferences though which are the following: (a) the fact that theoreticalrather than simulatedMSEs areconsidered,(b) we proposea two, rather

than one, parameter classification scheme, taking into account not only the demand intervalbut also the demand size variabilityand (c) we validate empiricallyour results. MSE is similarto the statisticalmeasureof the varianceof forecast errors (which consists of the variance of the estimatesproducedby the forecastingmethodunderconcern and the varianceof the actual demand)but is not quite the same since bias can also be explicitlyconsidered.MSE is a quadratic error measure and as such it may be unduly influencedby outliers but it is the only accuracymeasure that allows theoreticalresults to be tested in practice. In addition, MSE can be defined for all demand data series ratherthan ratio-scaleddata only, an issue that is obviously demand context. of great importancein an intermittent The lead time MSE assuming error auto-correlation In an inventory control application, the forecast errors producedby any estimationprocedureare usually assumed to be normal and independent(or Poisson for slow moving items). Usually the effect of non-normality is small. 'However, for lead times greater than 1 the errors will and this issue (the issue of autotypicallybe auto-correlated correlationof the forecast errors)has receivedvery limited attention'.6 By ignoring the auto-correlation term, we are most probably overstating the performance of the estimation inducedby proceduresunderconcernsinceauto-correlations bias or lumpinessare generallypositive. The MSE over a fixed lead time of duration L, assuming error auto-correlation, for unbiasedforecastingmethods is calculatedas follows:7 MSEL.T. =L{L Var(Estimates) + Var(Actual Demand)}

(1)

where the variance terms on the right-handside refer to a single period. For any biaseddemandestimationprocedure,it is easy to show that the lead-timeMSE is given by (2): =L { LVar(Estimates)+ LBias2 MSEL.T. + Var(ActualDemand) } It immediatelyfollows, from Equation (2), that the MSE over lead time L for method A is greaterthan the MSE over lead time L for method B if and only if + BiasA> Var(Estimates)B Var(Estimates)A + BiasB (3) where the subscripts refer to the forecasting methods employed. (2)

This content downloaded from 146.155.157.12 on Wed, 7 Aug 2013 14:51:14 PM All use subject to JSTOR Terms and Conditions

498 Journal oftheOperational Research Vol. 5 56, No. Society

The MSE of intermittent demand estimation procedures Crostoni showed that the performance of EWMA on intermittentdemand data series depends heavily on which points in time are considered. That is, if all the estimates producedby the method, at the end of everyforecastreview period, are taken into account, the method is unbiased whereasa certain bias should always be expectedwhen we isolate the estimates made only after a demand occurring period. The former scenario corresponds to a re-order interval(periodicreview)systemwhile the latterto a re-order level (continuousreview)system. Moreover,the varianceof the estimatesproducedby EWMA in those two cases is not the same. The samplingerror of the mean associated with this method is always higherwhen we considerall points in time. Regarding Croston's estimator (see Appendix A), the of the methodis independentof whichpoints in performance time are considered.Croston'smethod has been claimed to be unbiasedand of greatvalue to manufacturers forecasting intermittent demand.The method is widely used in industry and it is incorporatedin best-sellingforecasting(and stock control)softwarepackages.A ratherdisappointing empirical performanceof this estimator,though, has been reportedin the academicliterature8,9 when comparedwith conceptually such as EWMA or even simple moving simplerestimators, These have motivatedresearch, conducted averages. findings the first two authors of this by paper, on the theoretical properties of Croston's method. Syntetos and Boylan'o showed that Croston's estimator is biased. Based on Croston's approach and theoretical model (demand is assumed to occur as a Bernoulliprocess and thereforethe inter-demandintervals are geometrically distributed) the same authors" proposed a modification on Croston's method that is approximately unbiased Y I=21- )(4)

attentionto periodicreviewsystems,both 'all points in time' and 'issue points' should be considered.Nevertheless,only the formerscenariowill be exploredin this paper. Considering results presented in the academic literature1'10-12 the MSE over a fixed lead time L associatedwith EWMA, Croston's method and Syntetos and Boylan method (elsewherereferredto also as the 'Approximation' method3'"-13), when all points in time are considered, is calculatedas follows:
MSEEwMA=L L2

- p-1
P
(/2+

a2

[p2lP2

(5)

+ P 12 + ?]}
MSECROSTON -LL a

2-a

(p-1)

2 -

2-

+a2)+U + L 2_

PP

I2

P
p2

2 2+
P+

(6)

MSESYNTETOS &BOYLAN L{L

4(2-

[(p

31)

L 2L[
-a
[2 p2+2 p2

2 2
2 P

(7)

where p,' is the exponentially smoothed inter-demand interval,updated only if demand occurs in period t, z,' the exponentially smoothed size of demand, updated only if demand occurs in period t, a the common smoothing constant value used and Y,' the estimate of demandin unit time period t. The method'sperformance is independentof whichpoints in time are consideredand its theoreticalpropertiesare also discussedin AppendixA. For the purpose of illustratingour approach,we proceed to the case of the periodicreviewsystems(ie we by referring considerall points in time) since such systemsare most often used in practicefor the managementof intermittent demand At this point, we need to mention that issue point items.3'9 forecastscould also be relevantin the context of a periodic reviewsystemif stock is reviewedat the end of each periodin which a forecast is made. Hence, even if we restrict our

wherea is the common smoothingconstant value used, with P= 1-. In case of Croston's and Syntetos and Boylan method, the o valueis the smoothingconstantvalue used for the intervals. Following Croston's1 suggestion the same smoothing constant is also used for demand sizes although, following the suggestionof Schultz,14 a differentsmoothing constant could also be used. y and a2 are the mean and variance, respectively,of the demand sizes, when demand occurs, and p is the averageinter-demand intervalexpressed in numberof forecastreviewperiods (includingthe demand occurringperiod). Equation(5) and approximations (6) and (7) are based on the assumptionthat demand occurs as a Bernoulliprocess and thereforethe inter-demandintervals are geometrically distributed(see also Appendix A). Theoretical comparisons of the MSEs We start the analysisby assumingthat the approximatedor exact MSE of one method is greaterthan the approximated

This content downloaded from 146.155.157.12 on Wed, 7 Aug 2013 14:51:14 PM All use subject to JSTOR Terms and Conditions

499 ofdemand et al-Categorization AA patterns Syntetos

or exact MSE of one other method, and we try to specify underwhat conditionsthe inequalityunderconcernis valid. Consideringinequality(3) it is obvious that the comparison between any two estimationproceduresis only in terms of the bias and the variance of the one step ahead estimates associatedwith their application.That is, the length of the fixed lead time and the variance of demand itself do not affect the final results. At this point, it is importantto note that all the pairwise comparison results are generatedassuming that the same smoothing constant value is employed by all of the estimation procedures under concern. We recognize that the use of the same smoothingconstantmay put one or more but the issue methods at a relativeadvantage/disadvantage of sensitivityof the comparativeperformanceresultsto the application of the same a value has not been further explored. Comparisonresults (MSE Croston's method-MSE Syntetos and Boylan method) We firstcomparethe MSE of Croston'smethod with that of the Syntetos and Boylan method over a fixed lead time of length L (L > 1):
> MSESYNTETOS &BOYLAN MSECROSTON = (p-)4(p2)- p(~
"2

- 1)

-4)

(8)

2-2

for p>1,

1. Oo,

(All detailed derivationsof our results are available upon requestfrom the first author.) The theoreticalrule developedabove is expressedin terms of the squaredcoefficientof variation(CV2)and the average interval.The rule can then be furtheranalysed inter-demand (consideringdifferent possible values of the control parameters: L, y, p and a2) so that cut-off values can be determined. For any p> 1.32, inequality (8) holds (superior performance is theoreticallyexpected by the new method). For p< 1.32
BOYLAN

p= 1.32, therefore, Croston's method performs better for l<p<<1.32. The above resultsare valid for = 0.15 and approximately true for other realistico values (referto Table 1). From the above analysis, it is clear that there are four decision areas in which one method can be theoretically shown to performbetter than the other. For averageinterdemandintervalsand squaredcoefficientsof variationabove their correspondingcutoff values, we know with certainty which method is theoreticallyexpected to perform better. The same happens when either of the criteriatakes a value above its cutoff value, while the other takes a value below its cutoff value. The only areathat requiresfurtherexamination is the one formedwhen both criteriatake a value below their cutoff values. corresponding From Table 1 we can conclude that for all smoothing constant values that are likely to be appliedin practice,the new method should always perform better than Croston's method for anyp > 1.32unit time periodsand/or CV2>0.49. In the area that correspondsto p< 1.32 and CV2<0.49, shown to performbetter neithermethod can be theoretically in all cases. Our numericalanalysisresultedto assigningthis area of indecisionto Croston'smethod. (In the majorityof cases Croston'smethod performsbetter or if this is not the case the differences are so small that we are almost indifferentas to which estimator will be used.) Similarly, for the EWMA-Syntetos and Boylan method comparison, area will be assignedto the method that the corresponding in the other decision areas. Clearly, the worse performs derivationof a function, based on which the more accurate estimatorcan be exactlyidentified,would be welcomed,but such an exerciseis beyond the scope of this research. schemeis as indicatedbelow (Figure3). Ourcategorization demandcategoriesare as follows:area The corresponding area 2-lumpy; area not very intermittent); 1--erratic (but not 4-intermittent area 3-smooth; very erratic). (but Comparisonresults (MSE EWMA-MSE Syntetos and Boylan method) if MSEEWMA > MSESYNTETOS&BOYLAN
2

and only if

p(

1)[4p

12 forp>l, 0ocl1. Table1

(2 -

(2 - a) )2] )(2p2 p) 4p3

-

- (2

(9)

* If CV2> 0.48 then MSECROSTON > MSESYNTETOS&

* If then there is a p cutoff value (1 <p 1.32) CV2<0.48, below which Croston's method performs better (ie the inequalityis not valid). For example if CV2 =0.15, then the cutoff value is p= 1.20. For 1<p 1.20, Croston's method performsbetterand for p> 1.20 the Syntetosand Boylan method performsbetter. As the ratio CV2 decreases, the p cutoff value increases, and for CV2=0.001 the cutoff value is method-MSESyntetosand Boylan MSE Croston's method

a Smoothing constant value p-Cutoff value CV2 cutoff value

0.05 0.10 0.15 0.20 1.32 1.32 1.32 1.31 0.49 0.49 0.48 0.47

This content downloaded from 146.155.157.12 on Wed, 7 Aug 2013 14:51:14 PM All use subject to JSTOR Terms and Conditions

500 Journal oftheOperational Research Vol. 5 56, No. Society

p = 1.32(cut-off value)

Overall comparisonresults When all points in time are considered, the Syntetos and Boylan method performsbetter than the other two methods for p > 1.32 and/or CV2> 0.49. For p < 1.32 and CV2<0.49, Croston'smethod is theoreticallyexpectedto performbetter than the othermethods.This resultis, at least intuitively,not what one may expectand it is attributedto the largevariance associatedwith the estimatesproducedby EWMA. Simulation exercise

& Boylan Syntetos 1

& Boylan Syntetos 2

CV2= 0.49 (cut-off value) Croston 3 & Boylan Syntetos 4

Figure 3 Cutoff values (Croston's method-Syntetosand Boylanmethod). An analysis similar to that presented in the previous sub-section reveals that for smoothing constant values that are commonly applied in practice, the Syntetos and Boylan method should always perform better than the EWMA method for any p > 1.20 unit time periods and/or CV2>0.49. Comparisonresults (MSE EWMA-MSE Croston's method)
> MSECROSTON MSEEWMA

if and only if
U2> 1- p 0-2 2 P (10)

But
a2

2>

1 p

for p>l1, 0<

1.

This is indeed an unexpected result if we consider the unbiased nature of EWMA when all points in time are considered.The theoreticalexplanationis the high sampling error of the mean of EWMA, but the result may also be explainedin terms of the approximatenature of Croston's MSE. The results of the above-discussed pairwise comparisons indicate that if all points in time are considered,then one of the estimation procedures specificallydesigned to deal with intermittencecould be utilized for all SKUs. For fast demand items, the Syntetos and Boylan method's performance is in general though inferior to that of Croston's method. The opposite is the case when more intermittent and/or more irregular demand patterns are considered.

A sample of 3000 files is available for the purpose of this research. Each file consists of 24 demand time periods/ months. The averageinter-demand intervalrangesfrom 1.04 to 2 months and the average demand per unit time period from 0.5 to 120. The average demand size, when demand occurs, is between 1 and 194 units and the varianceof the demand sizes between0 and 49 612 (the squaredcoefficient of variation ranges from 0 to 14). That is the sample contains: slow movers; intermittentdemand series with a constant, or approximatelyconstant, size of demand; and very lumpy demand series. What follows is the number of files that falls within each of the four classes of demandthat result from our classificationscheme: erratic (but not very intermittent)-441 files; lumpy-314 files; smooth-1271 files; intermittent(but not very erratic)-974 files. To initializea particularmethod'sapplication,the first 13period demand data are used. Therefore,the out-of-sample comparisonresultsthat are reportedin this paper will refer to the latest 11 monthly demand data. The first EWMA estimate is taken to be the averagedemand over the first 13 periods. In a similarway, the first exponentiallysmoothed estimate of demand size and inter-demandinterval can be based on the averagecorrespondingvalues over the first 13 periods. If no demand occurs in the first 13 periods, the initial EWMA estimateis set to zero and the inter-demand interval estimate to 13. As far as the demand size is concerned,it is more reasonableto assign an initial estimate of 1 ratherthan 0. Optimizationof the smoothing constant values used was not consideredbecause of the very few (if any) demand occurringperiods in the data sub-setwithheld for initialization purposes. Since the available data series were so short (24 periods),the initializationeffect is carried forward by all estimators on all out-of-sample point estimates. Clearly, longer demand data series would have been welcomed. However, longer histories of data are not necessarily available in real-world applications. Decisions often need to be made consideringsamplessimilarto the one used for this research. The smoothing constant value is introducedas a control parameter.In an intermittentdemandcontext, low smoothing constant values are recommended in the literature. Smoothingconstantvalues in the range 0.05-0.2 are viewed as realistic.1'4'8 From an empirical perspective, this range

This content downloaded from 146.155.157.12 on Wed, 7 Aug 2013 14:51:14 PM All use subject to JSTOR Terms and Conditions

501 ofdemand eta/l--Categorization AA patterns Syntetos

covers the usual values of practicalusage. In this research four values are simulated:0.05, 0.10, 0.15 and 0.20. The lead time is also introducedas a control parameter. The lead times consideredare 1, 3 and 5 periods.

MSE results A Z2-testwas decided to be the most appropriateway of assessingthe validityof the theoreticalrules proposedin the intervaland the previoussection.The averageinter-demand squared coefficient of variation of the demand sizes were generatedfor all 3000 files and the proposed categorization rules were used in order to indicate which method is theoreticallyexpectedto performbetter or best in each one of those files. The null hypothesis developed is that the performance of the methods is independent of what is expectedfrom the theory. Croston'smethod is always(on everydemanddata series) expected to perform better than EWMA, independentlyof As such, the Z-test the demand data series characteristics. statistic for the population proportion will be used to test whether or not the number of files on which Croston's method out-performsEWMA is significantlygreater from the numberof files on which EWMA performsbetter. For testing the categorizationrule regardingall methods' performance a 3 x 2 (2 degrees of freedom) or 4 x 2 (3 degreesof freedom)table will be used dependingon whether or not ties occur. The Z2-test statistic values are indicated, for different smoothing constant values, in Table 2. In brackets we present the number of files that the methods perform as expected. The shaded area refers to the Croston-EWMA pairwisecomparisonand the valuespresentedare the Z-test statistic values for the population proportion. Statistically significant results at 1% level are emboldened while significanceat 5% level is presentedin italics.

The results generated on MSEs indicate the practical validityof the rulesproposedin this paper.The performance of the methods is clearly not independent of what the categorization rules suggest. This is true at a pairwise comparison level and when the overall rule (regardingall methods' performance)is considered. The results, though regardingthe EWMA-Croston comparison,are not what for the 'smoother'SKUs. one may have expectedempirically There is some empirical evidence to attribute that differenceto the high variabilityof the errorsproducedby the EWMA estimator.'2The decompositionof the empirical MSE into its constituentcomponents(bias squared+ variavariabilityof demand+ variabilityof the estimates) would obviouslyenable the rigorousassessmentof the contribution of these componentsto the empiricalMSE. Nevertheless,no such resultshave been generatedin our empiricalstudy and this issue requiresfurthersimulationson real data.

Conclusions demandpatternsfacilitates The categorizationof alternative the selection of a forecastingmethod and it is an essential element of many inventory control software packages. Despite the importanceof this issue though, the problem of categorizingdemand patterns has received very limited attention so far in the academicliterature.Some work has been performedin this area, which neverthelesslacks either universalapplicabilityor empiricalvalidationof the results. The common practice in inventory control software industry is to arbitrarilycategorize demand patterns and then select an estimation procedure in order to forecast future requirementsand manage stock efficiently. Ultimately,the objectiveis the selectionof a forecastingmethod. Based on Johnston and Boylan4we propose an alternative approachto the categorizationproblemaccordingto which direct comparison of the forecasting methods results in specifyingthe demandcategories.The pairwisecomparisons

Table2

test results X2
Croston-S.B.a 15.77 (1631) 13.18 (1624) 20.75 (1653) 17.22 (1640) 13.02 (1625) 30.56 (1688) 28.27 (1678) 16.79 (1644) 31.96 (1694) 18.70 (1654) Overall categorizationrules

a
0.05
0.1 0.15

L.T.

Croston-EWMA 5.29 (1645) 4.35 (1619) 3.83 (1605) 10.11 (1777) 9.02 (1747) 6.83 (1687) 13.04 (1857) 13.04 (1857) 14.83 (1906) 15.12 (1914)

EWMA-S.B.a 1.10 (1691) 4.41 (1726) 6.96 (1709) 17.98 (1827) 19.02 (1826) 17.34 (1753) 53.58 (1925) 44.66 (1913) 84.10 (1991) 60.77 (1949)

1
3 5 1 3 5 1 3 5

0.33 (1509)

0.84 (1633)

16.45(1634)

9.19 (979)
12.80 (1111) 20.47 (1160) 1.60 (1081) 11.94 (1238) 17.43 (1271) 6.20 (1160) 13.05 (1339) 22.20 (1364) 20.86 (1405) 21.37 (1418)

0.2

1
3 5

8.47 (1732)

28.99 (1802)

53.58 (1744)

16.22(1230)

and Boylanmethod. aS.B.standsfor Syntetos

This content downloaded from 146.155.157.12 on Wed, 7 Aug 2013 14:51:14 PM All use subject to JSTOR Terms and Conditions

502 Journal oftheOperational Research Vol. 5 56, No. Society

are based on the theoreticalMSEs and they can indicate universally valid regions of expected superiority of one method over the other. We illustrate our approach by consideringCroston'smethod, Syntetosand Boylanmethod and EWMA. The validityof the rulesproposedis confirmed demand by means of simulation on 3000 real-intermittent data series.The empiricalresults also demonstratethe lack of sensitivityof the cutoff points proposedto the smoothing constant value being used. The resulting categories of any demand classification scheme for inventorymanagementare ultimatelymeant to serve the inventoryobjectivesof improvedcustomerservice levels and/or reductionof stock holdings.In that respectthe its bias, forecastaccuracyof an estimator(and subsequently or the lack of it, and its samplingerrorof the mean) is not the only theoretical and practical concern regarding That is to say, specificdemanddistributional categorization. as assumptions well as certain stock control models should also, eventually,be considered.As far as the formerissue is concerned,it is importantto note that an interestingavenue for furtherresearchcould also be the comparisonof nonparametric(bootstrapping)versus parametricapproaches, the latter comprisinga forecastingmethod and a particular The coefficientof variationof assumeddemanddistribution. demand sizes has been shown in this paper to be very importantfrom a forecastingperspective.Nevertheless,we should mention that the use of this criterionalso impliesthe non-complianceof the true demand size distributionwith standardtheoreticaldistributions(such as the geometricor the logarithmic).As the coefficientof variation rises, so it becomes more difficult to justify the use of any standard distributionand that, as pointed out before, opens up a should underwhat circumstances furtheravenueof research: a distributionalapproach be used at all. What has been presentedin this paper can be perceivedas only a first step towardsa more systematicand meaningfulapproachto the categorizationproblem. References
1 Croston JD (1972). Forecasting and stock control for intermittent demands.Opl Res Q 23: 289-304. 2 Williams TM (1984). Stock control with sporadic and slowmoving demand.J Opl Res Soc 35: 939-948. 3 Eaves AHC (2002). Forecastingfor the orderingand stock holding of consumablespare parts. PhD thesis, Lancaster University,UK. 4 JohnstonFR and Boylan JE (1996). Forecastingfor items with demand.J Opl Res Soc 47: 113-121. intermittent 5 SegerstedtA (1994). Inventory control with variation in lead times, especially when demand is intermittent. Int J Prod Econom35: 365-372. 6 Fildes R and BeardC (1991).Productionand inventorycontrol. IJOPM 12: 4-27. 7 Strijbosch LWG, Heuts RMJ and van der Schoot EHM (2000). A combined forecast-inventorycontrol procedure for spare parts. J Opl Res Soc 51: 1184-1192.

8 Willemain TR, Smart CN, Shockor JH and DeSautels PA a (1994). Forecastingintermittentdemand in manufacturing: comparativeevaluationof Croston'smethod. Int J Forecast10: inventory control and demand forecasting methods for low demanditems. J Opl Res Soc 48: 700-713.

529-538. 9 Sani B and Kingsman BG (1997).Selecting the bestperiodic

AA andBoylan 10 Syntetos JE (2001). On thebiasof intermittent

demandestimates.Int J Prod Econom71: 457-466. 11 Syntetos AA and Boylan JE (1999). Correcting the bias in

forecastsof intermittent demand. Presentedat the 19th June1999, International on Forecasting, Symposium Washing-

ton DC, USA. 12 Syntetos AA (2001). Forecastingof intermittent demand.PhD thesis, Brunel University-BuckinghamshireChilternsUniver13 Eaves AHC and Kingsman BG (2004). Forecasting for the orderingand stock-holdingof spare parts. J Opl Res Soc 55: 431-437. 14 Schultz CR (1987). Forecasting and inventory control for sporadic demand under periodic review. J Opl Res Soc 38: 453-458. 15 Rao AV (1973). A comment on: forecastingand stock control

UK. sityCollege,

for intermittent demands. OplRes Q 24:639-640.

Appendix A: Croston's method, Syntetos and Boylan method

Croston,' as corrected by Rao,'5 proved the inappropriateness of exponential smoothing and he also expressed, in a

quantitativeform, the bias associated with the use of this method, when dealing with intermittentdemands.
Moreover, assuming a stochastic model of arrival and size

of demand, Croston introduced a new method for characterisingthe demandper period by modellingthe demand in one period from constituentevents. Accordingto his method, separateexponentialsmoothing estimatesof the averagesize of the demandand the average intervalbetweendemandincidencesare made after demand occurs.If no demandoccursthe estimatesremainexactlythe
same.

Let: 1/p,be the Bernoulliprobabilityof demandoccurring in period t; pt the geometricallydistributed(including the first success, ie demand occurring period) inter-demand interval; p,' the exponentially smoothed inter-demand interval, updated only if demand occurs in period t, E(p,)= E(pt') =p; z, the normally distributeddemand size, when demand occurs;z,' the exponentiallysmoothed size of demand, updated only if demand occurs in period t, the common smoothing constant value E(z,)= E(zt')= i; o~
used and Y, the demand in unit time period t.

Under these conditionsthe expecteddemandper unit time period is E(YI,)=t/p. Following Croston's estimation procedure,the forecast, Y, for the next time period is given by: Y,' z,'/p,' and, according to Croston, the expected estimate of demand per period in that case would be:
E( Y,') =E(zt'/p,')= E(z,')/E(p,')= t/p unbiased). (ie the method is

This content downloaded from 146.155.157.12 on Wed, 7 Aug 2013 14:51:14 PM All use subject to JSTOR Terms and Conditions

ofdemand 503 AA et al--Categorization patterns Syntetos

The varianceof the forecastsproducedby this method is calculatedby Croston as follows:

Var
t42-a

following: (A.4) Y-(=1The expected estimate of demand per unit time period as well as the variance of the estimates produced by this method are given by (A.5) and (A.6), respectively
E(Yt )E ((I -2 P 2p2 (A.5)

zt

(A.1)

p
+-

Syntetos and Boylan" showed that Croston's method is biased and that the expected estimate of demand per unit time period is not as calculatedby Croston, but rather

(p-)

(A.2) (A.2)

Moreover,the samplingerrorof the mean was also found to be different:

Var( Y)

--Var-

(1

Var

(A.6)

t)I

a -

P(P - 1)
p4

2
22

02

2 The researchers 2-Va proposed a new estimator that takes into 2-a p2-,account the smoothingconstantvalue used and which is the

zl

Received November 2003; accepted June 2004 after one revision

This content downloaded from 146.155.157.12 on Wed, 7 Aug 2013 14:51:14 PM All use subject to JSTOR Terms and Conditions