Sie sind auf Seite 1von 17

Demand forecasting and sharing strategies to reduce fluctuations and the bullwhip effect in

supply chains
Author(s): Y Barlas and B Gunduz
Source: The Journal of the Operational Research Society, Vol. 62, No. 3, Special Issue: Supply
Chain Forecasting and Planning (March 2011), pp. 458-473
Published by: Palgrave Macmillan Journals on behalf of the Operational Research Society
Stable URL: http://www.jstor.org/stable/41058923
Accessed: 08-01-2016 15:25 UTC

REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/41058923?seq=1&cid=pdf-reference#references_tab_contents

You may need to log in to JSTOR to access the linked references.

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.

Operational Research Society and Palgrave Macmillan Journals 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 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
Journalof the Operational Research Society (201 1) 62, 458-473 © 201 I OperationalResearchSocietyLtd.All rightsreserved.0160-5682/1
I *^>-

wvwv.palgrave-journals.com/jors/

Demandforecasting to
and sharingstrategies
and thebullwhipeffect
reducefluctuations in
supplychains
Y Barlas1*and B Gunduz2
Istanbul,Turkey;and Ford Otosan,Golcuk,Kocaeli,Turkey
Bogazici University,
Supplychaininventories are proneto fluctuationsand instability.
Known as thebullwhipeffect, small
variationsin theenditemdemandcreateoscillationsthatamplify throughoutthechain.Byusingsystem
dynamicssimulation, we investigate
someofthestructural sourcesofthebullwhipeffect,and explorethe
effectiveness of information sharingto eliminatethe undesirablefluctuations. Extensivesimulation
analysisis carriedout on parametersof somestandardorderingpolicies,as wellas externaldemandand
lead-timeparameters. Simulationresultsshowthat(i) a majorstructural cause of thebullwhipeffectis
isolateddemandforecasting performed at each echelonofthesupplychain,and (ii) demandand forecast
sharingstrategiescan significantly reduce the bullwhipeffect,even thoughtheycannot completely
eliminateit. We specifically
showhow each policyis improvedby demandand forecastsharing.Future
researchinvolvesmore advanced orderingand forecasting methods,modellingof otherwell-known
sourcesof bullwhip,and morecomplexsupplynetworkstructures.
JournaloftheOperational ResearchSociety(2011)62, 458-473.doi:10.1057/jors.2010.188

Keywords: supplychain; bullwhipeffect;demandforecasting;


information
sharing;systemdynamics;
simulation

Introduction demand informationcould significantly reduce the


bullwhip effect.Xu et al (2001) and Lee and Whang
Supplychain inventories are prone to fluctuations and
that sharingof the demand forecastand
Small changesin the end item demand can (1998) report
instability.
createinventory and orderoscillations thatamplify as one inventoryinformationis effectivein reducingorder
fluctuationsand safetystocks. Gavirneniet al (1999)
movesup in thesupplychain(Forrester, 1961,Chapter12;
compares the no-information-sharing case against two
Sterman,1989; Sterman,2000, Chapter 17, p 18). This different of
types information-sharing policiesused by the
phenomenonof amplification of oscillationsthroughthe
retailer(partialand completesharing)in a simpleone-
supplychainis also knownas thebullwhip effect(Lee et al, chain. Gallego and Özer (2001)
retailer-one-supplier
1997;Chen et al 1998;Xu et al, 2001). searchesoptimalpoliciesfor withand withoutdemand
Lee et al (1997) identifiesfour main causes of the
information-sharing cases in a two-stagesupplychain,
bullwhip effectas: demand signal processing,order where the retailerbatches orders and faces Poisson
batching,rationinggame, and price variations.Chen demands.Cheng and Wu (2005) show how information
et al (1998) argues that the bullwhipeffectis due, in
sharingcan reduceinventory costsin a two-level chainwith
part,to theneed to forecastthedemand.Sterman(2000, retailers. et al show that
multiple Dejonckheere (2004)
Chapter 17 and 18) and Forrester(1961, Chapter 12) information is if not
sharing very beneficial, indispensable
show that delays inherentwithin the supply chain
in order-up-to-Spolicies since the magnitudeof the
togetherwith demand forecastingand distortioncan reducedat higherlevels
createamplifiedoscillations. bullwhipcan thusbe significantly
in thechain.However,theynotethatinformation sharing
Supplychainliterature and management practicefocus cannot eliminate the and
on coordinationpolicies based on information completely bullwhip. Jeong
sharing discusses the of a multi-echelon
Maday (1996) stability
among supply chain membersin order to reduce the chainfroma feedbackcontroltheoretic
Chen et al (1998) arguesthatcentralizing supply perspective.
bullwhipeffect. Silveret al (1998) suggestsdemandsharingand echelon
* inventory policyimplementations. Authorspropose that
Correspondence:Y Barlas, Bogazici University,Departmentof
Industrial 34342,Bebek,Istanbul,Turkey.
Engineering,
each stage apply echelon (s, S) policy in whichan agent
E-mail:ybarlas@boun.edu.tr monitorsitstotalecheloninventory level.Chenet al (2000)

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
- Demand
Y BariasandB Gunduz andsharing
forecasting to reducefluctuations
strategies andthebullwhip insupply
effect chains 459

demonstratethe fact that smootherdemand forecasts used in systemdynamicsmodels,is normally


ruletypically
reducethebullwhipeffect, and longerlead timesincreaseit. time-continuous.As will be explained below, in the
They also show that for both moving average and assumedparametersettings, it was possibleto represent
exponentialsmoothingforecasts,the veryinclusionand the anchor-and-adjustorderingrulewithDT = 1 as well,
need for estimationof a linear trendparameterinto withoutcausinganyerroneousdynamics.
the forecastingmodel results in increased bullwhip. The basicgenericequationsofthemodelaredescribed in
Dejonckheereet al (2002) analyzetheeffects of constant, thissection(exceptthe policy-specificorderingequations
linear,and quadraticexponentialsmoothingalgorithms thatare presentedlaterin separatesections).
on thebullwhip.Theyshowthatthebullwhipemanating Local inventory(LI) increases with arrivals and
fromthetrenddetectionalgorithms (linearand quadratic decreaseswithshipments:
or exponentialsmoothing)are reduced by loweringthe U- , = U-,r-i + (Ai,t- Siit) [Goods] (1)
exponentialsmoothing in
constantused thesealgorithms.
Morerecently, Datta et al (2007)analyzestherelationships whereAitis thearrivalsto stage/and Sittis theshipments
betweendemandand orderforecasting and the bullwhip fromstage/in periodt.
effect,and proposes an advanced forecastingmodel In transitinventory, the goods shippedby the upper
stagethathave not yetarrived:
(GARCH) forsupplychainmanagement.
The purposeof our researchis twofold:(1) to under- -
ITiit= ITUt-x+ (Si+U Aitt) [Goods] (2)
standsome of the underlying structures and factorsthat
fluctuations and the bullwhipeffect Goods in transitarriveafteran exponential(gradual)
generateinventory
delaystructure:
throughthesupplychain;and (2) to exploretheeffective-
ness of some management strategiesin ameliorating this Ai,t = ITi,t/LTi [Goods/Period] (3)
undesirablebehaviour.We particularly focuson uncoor-
whereLT¡ is thetransitlead timeneededforshipments by
dinated demand forecastingas a major cause of the to reach i.
stage(i+l) stage
bullwhipeffect,and sharingof demand and forecast (SR) fora stage/is the sum of
Shipmentrequirement
information as a counter-bullwhip managementpolicy. demandfaced(Dit) at timet and backloggedorders(BLit):
SystemDynamics modellingis used as the research
platform. SRi,t = BLUt+ A, t [Goods/Period] (4)
If thereis enoughLI, the requiredamountis shipped
immediately portionof
in one period.Ifnot,theunfulfilled
The model structure ordersis added to BL:

We considera three-stage supplychain systemconsisting Siit= mm(SRiìtì


Llht) [Goods/Period] (5)
of identicalagentswhereeach agentordersonlyfromits -
BLi,t+x= BL¿it+ A, t Si,t [Goods] (6)
upper agent (/+1). An agent ships goods immediately
upon receivingthe order,if thereis sufficient on-hand Net inventory (NI) is theLI afterthebackloggedorders
inventory.Orders be
may partially and
fulfilled, unfulfilled are subtracted:
orders are backlogged.Shipped goods arrive after a t = LU,t - BLi,t [Goods]
JV7/, (7)
constanttransitlead time(LT). The model represents a
set- Agents are assumed to be unaware of theexactdemand
general un-capacitatedproducer-wholesaler-retailer
ting.The uppermoststage(producer)places ordersto an patterntheyare facing,so theymustforecastthe future
unlimitedsource, so there is no backloggingof the demand. Simple exponentialsmoothingis used as the
producer orders by the raw material/parts supplier. forecastingmechanism.Thus, the expecteddemand is
Alternatively,we can thinkof a factorythatalwayskeeps calculatedby:
enoughstockof raw materialand parts.This assumption Eiit=£,,,_!+ (l/EATi)(A,r-i- Eiit-')
is made in a senseto draw a practicalmodel boundary,
elsethesameinventory structurewouldhave [Goods/Period] (8)
management
to be cascaded too many times,withoutadding any whereE ATi is the expectationadjustmenttimeused by
conceptualor novel dimensionto the research.(See the stage /. Dht_xis the demandfaced by stage /. The end
stock-flowdiagramsand equationsbelow.) demandDx is an externalinputto be describedin thenext
For consistencywith models and policies in the sectionand demandsD2 and D3 facedby stagetwo and
inventory timeis modelleddiscrete(DT = 1 time
literature, threeare actuallytheordersplaced by theirlowerstages.
period). This makes the model time-discrete that is (Di = 0¿_x, whereOt are to be describedin the following
necessaryto representstandard orderingpolicies like sections.)
Order-up-to-S and (s, S), as will be describedbelow. For any agent,the total expecteddemandduringLT
Anotherpolicyanalyzed,the anchor-and-adjust ordering (DiL) is simplycalculatedby theexpecteddemandforone

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
460 Journal Research
oftheOperational Vol.62,No.3
Society

by LT:
periodmultiplied {ITi t+ BLi+l> ,). Thus, the above formulationdoes not
implyan unrealistic assumptionin a realapplication.
A,l= {LTi){Eiit) [Goods] (9) The stock-flow diagramsfor retailer,wholesaler,and
producerare shownin Figure1 and 2.
Each agentcontinuously monitorsher inventory posi-
tion(IP), definedas the sum of her local inventory
(LI),
The demandpattern
goods(IT) fromtheupperstage,backlog(BL) at
in transit
theupperstage,minusBL at herown stage: Three typesof demandinputare used: (1) deterministic
'step up and down' demand for testing/verification
IP i = Lit + IT i + BLM - BLi [Goods] (10)
purposes;(2) iid randomdemand(Normal(20,2)) in some
In reallife,itis difficult foran agentto knowhowmuch selectedcomparisonruns; and (3) stationaryautocorre-
ofwhite
of its ordersare in transitand how muchare backlogged lateddemand(obtainedbyexponential smoothing
noise, withan autocorrelationperiodof five)used in most
by theupperstage.So, theexplicitinclusionof the term
runs in this paper, unless otherwise noted. (Figure 3
'ordersbackloggedby theupperstage(BLi+1 ,)' in theIP
illustrates
these threedemand patterns.)
formulation mayseemtoo idealistic. However,all ordering
policies tested in this researchutilizethesumof in-transit
goods and upper stage backlogs{ITt+ BLi+x), whichis
the difference betweenthe total ordersgivenand Orderingpolicies
simply
the ones that have arrived.An agentdoes not need to Threebasicinventory management policiesaretestedwithin
obtain specificBL information fromits upper stage in the supplychainmodel:Order-up-to-S policy,a standard
orderto determine totalordersthathave not yetarrived system dynamics ordering policy (anchor-and-adjust

rp £J Stage1: Retailer XX {}

/hipment2 In Transit N. ^'


^^^^^^^y^ne''men'or^

supplierdispatch ^T arrival T shipment '. '

transitlead time / T ^v X

/ ' / '
requirementX^erBackb/
Order 2
Backlog / ' _
y/ 4^

| ®- -^i)^-. - /backlog
i ;
?<s --~~~"*l J^K^^
^^^^--^__^^/ ""
_^--^-^^^^^^ ' / change
position '
inventory /~~~" >v I

~~
( Expected Demand
p ~~~~~^~^""^^^ ^^^-^n^/
jf^S. 'JVw^ ^^^í M ^^c^^kC^i demand
orderdocisiorrv N K.J ^Cw
'^ orderup to level c ' V
^'^^^ ^^^ change
expectations

inventoty
expectation
adjustmenttime
adjustmenttime

Figure1 Generalstock-flow
diagram model:stageone,retailer.
ofthesupply-chain

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
Y ßarlas - Demand
andß Gunduz andsharing
forecasting toreduce
strategies fluctuations
andthebullwhip insupply
effect chains 461

Ç0 Tlj 5fs§€2 Wnojesa-*" ft

f shipment
3 L«al Inventory 2 X.
'n Transit ' 2^>^i5st trtvertory

dispatch2
supplier ^T ? ^^^-^wimnt2 ^v I
arrival1!

OrderBacklog3 / transit
teastime2 / . jl^^X X
* smpmtnt
r / I / X^Qréer Backlog^ Nt

- ^"^ '
mvenwry
pos-tt^ té^' ^oacklog
'^/ Jchange2
y ^^^'

Q*^~^ -----w^^^^ ^/ ExpectedDermrjH J^^^^J

X. orderypto leve 2 /
^ expectation
'v ^s^ ' cliarì^2

^^ ¡
inventory '

C© T3JJ Stage2 ^oda^* ||

InTransjfi
„ ^ '^
/-°*^'
3
3 ...--'netinventory
LocalInventory n.

'- ^- 1
J f -5hjpment3'
5Uppl^<ispat*3 arriva^
/ ' / / ^'S >v J

/ I transit
ieai time3 / ir'. X
I I A / shipment'i OriSer
Backtoft^'
ji ' 3 c !
j ^/ res|uirerkent ^^ V-.

I 1
inv^oiry
posibrT?^^^ J^--^i /^^
X /* ^*S"*^~*~"*^^^ '. I chan^3

(D*^~^~^^^ _ Demand
Expected 3 J^^V^
O^" "~^^^i ^>mZ
orcteroï^
' orderup to level3
eC^É^SÇ"
^ /"*V/~ ^UJ
t
X,^ ^-'"^ » / expectation
A "^^^-^__ - -- ~^*^ '^an§e3 ^
inventory oréerVision 2
time3 ex^tatbn
expectation
ad|iÄtment
adjustmenttime3

Figure2 Stock-flow
diagramof stagetwo and three:wholesalerand producer.

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
462 oftheOperational
Journal Research Vol.62,No.3
Society

Figure3 Threedifferent
demandpatterns
usedin simulation
experiments.

policy),and (s,S) policy.All stagesare assumedidentical, variation.Thus,theformulais:


so foreach testedpolicyall stagesuse thesamepolicywith
thesameparameters. In reference runs,each stagelocally £. , = (LT/+ Ki)EUt [Goods] (12)
decides on the order quantitywithoutconsideringthe where
LT¿ is thetransitLT forstage/,K, is thelead time
overallsupplychain. inflationconstant',and Eht is the expecteddemand(or
expectedordersfromstage i-') estimatedby stage /.
Order-up-to-S policy (Recall that DLL= (LTl)(Eht) anyway,so that the term
(Kt)Eitis theextraadditionto DL to accountforrandom
Order-up-to-S policyis thewell-known base-stockpolicy
variation.)
wherean agentordersthequantityneededto bringits IP In thispolicy,the order-up-to-level S is updatedeach
up to a base-stocklevelS, whenever it fallsbelowS. The
period,as theexpecteddemandis updated.The resulting
associatedordering equationis: inventory dynamics (with autocorrelated demand),
= -
Oit max((S,-, IPi,t)/IAThO) . LT¡ = 3, Kj = 2 is seen in Figure4. Note the oscillations
({ }
[Goods/Period] and the bullwhipeffect(amplification) along the supply
chain, as one moves from the retail end toward the
WhereOh, is theorderdecision,Sht is theorder-up-to- as
producer.The increasein the bands of fluctuations
level, IPj, t is the inventoryposition,and IAT¿ is the we move up in the supplychain is graphically indicated
inventory adjustment time.In thestandardorder-up-to-S more
clearlyin Figure5.
policy,thediscrepancy is immediately ordered,so we set
inventory adjustment time(IAT) to one.
The order-up-to-level (Sh t) can be setin differentways. Systemdynamicsanchor-and-adjustpolicy
One approachis to computeit byS = DL + koL,whereDL
is expecteddemandduringLT, koLis safetystockwhereoL Thisis theanchor-and-adjust policywidelyusedin System
is thestandarddeviationof forecasterrorsof LT demand, Dynamicsliterature (Sterman,2000; Barlas, 2002). This
LI at a desiredlevel.The associated
and k is a constantselectedaccordingto desiredservice policytriesto stabilize
level.In anycase,theidea is to setS at a levelgreaterthan orderequation is the following:
DL to accountfordemandvariation(See Gündüz,2003). Oit =max((/*,- Iiit)/IATi + (SL*, - SLiit)/SLAT¡
Anotherpracticalformulasuggested forS is to 'inflate'the
shipmentLT by a factorK, to account for demand +£,-,„0) [Goods/Period] (13)

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
Y BariasandB Gunduz-
Demand andsharing
forecasting to reducefluctuations
strategies andthebullwhip
effect
insupply
chains 463

Figure4 Netinventories
whenorder-up- 4: producer).
to-Spolicyis applied(2: retail;3:wholesaler;

5 Thebullwhip
Figure thebandsofinventory
effect: fluctuations
areamplified
as onemovesfromtheretailendto thewholesaler,
andthento theproducer
end(markedbyA, B, andC respectively).

whereT*htis the desiredinventorylevel and SL*,t is t and SLh t represent


//, leveland supplyline
inventory
the desired supply line, IAT and SLAT are inven- thatare definedas follows:
toryadjustmenttime and supplyline adjustmenttime,
respectively. llt = LIit-BLlt [Goods] (14)

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
464 Journal Research
oftheOperational Vol.62,No.3
Society

SLi,, = ITi,t+ BLi+i,t [Goods] ( 15) graphsdue to space limitation(the readeris referred to
Gündüz,2003).
The desiredlevels of inventoryand supply line are Finallynote that,as explainedbefore,forcomparison
determined
typically by: and consistency betweenall policies,theanchor-and-adjust
SL* t = LTiX Et,t [Goods] policy runs reportedin thisarticleare done withDT='
( 16)
(ie with a discretemodel).Since systemdynamicsmodels
are typicallycontinuous,we have also testedthe model
I*t = M [Goods] (17)
withseveralsmaller¿lvalues to makesurethattheobtai-
whereM is somechosenconstant. nedbehaviours arenotsensitive to DT. Indeed,verysimilar
Note thatin thispolicy,the desiredsupplyline SL*,t is inventory and
oscillations bullwhipeffects wereobtained
adjustedaccordingto theexpecteddemand,so as to yield withsmallerDT values,provingthatthe resultsare not
an arrivalrate equal to the expecteddemand at equili- sensitiveto DT='. (We are unableto providethegraphs
brium.The desiredinventory level on the otherhand is pleasesee Gündüz,2003.)
due to space restrictions;
takenas constantin thesimplestversionof thepolicy.
Inventorybehaviourin thiscase (withautocorrelated
demand,M= 100,L7=3, and lAT^SLAT^ 1) is seen (s,S) Policy
in Figure6. Justlike withthe order-up-to-levelS policy,
(s,S) Policy is a reviewpolicywhereordersare placed
we observeoscillationsand thebullwhipeffect(amplifica- IP dropsto the
S, whenever
to raiseIP to order-up-to-level
tion) throughthe supplychain,fromthe retailerto the reorderpoints or below.The orderequationis as follows:
wholesaler and thento theproducer.(ComparingFigures4
and 6, also notethatthebehaviourpatternsof ordersand Oit = (S-IPitt)/IAT ifIPi,t<:S
inventorieswith order-up-to-S policy and anchor-and- Oi,t = 0 otherwise [Goods/Period] (18)
adjust systemdynamicspolicyare quite similar.)As an
alternative,thedesiredinventory in theadjustment equa- In thestandard(s,S) policy,thediscrepancyis immedi-
tionabove can be definedas proportional to theexpected atelyordered,so we setIAT to one. The parameterss and
demand,namelymEht.With thisformulation, both the S musttypicallybe determined as functionsof LT and
oscillationamplitudesand thebullwhipeffect increaseas a theexpecteddemandE. In general,thelargerLT and E,
resultof stronger(double) effectof theexpecteddemand the highermustbe the reorderpoints, due to a riskier
on orders.We are unable to providethe corresponding the largerE, the highermustbe the
situation.Similarly,

when anchorand adjust policyis applied.


Figure6 Net inventories

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
Y BariasandB Gunduz-
Demand andsharing
forecasting to reducefluctuations
strategies andthebullwhip
effect
insupply
chains 465

S in orderto preventfrequent
order-up-to-level shortages. of thebullwhipeffect.In (s,S) policy,ordersare in effect
Thus, even if we assumeconstantLT, s and S mustbe batched.
updatedin practice,sincethe demand(henceEt ) would A relatedissuewouldbe thefrequency ofupdatings and S.
be varyingin time.In lightof this,thereorderpointsht> The above formulasassume that s and S are updated
and theorder-up-to-levelSht forstage/at timet, can be in everyperiodt.It is likelythatin somesituations,
s and S
computedby thefollowingequations: are updatedless frequently, if
especially thedemandtends
to changeslowly.Anotherextreme wouldbe notto update
siit= (LTi)(Eiit) + SSi [Goods] (19) s and S at all through theentiresimulationhorizon.In this
articlewe also presentthisextremecase of keepings and
Si,t = siit+ q £/,t [Goods] (20)
S constantthroughthe entirehorizon,in the Policy
whereSSi is some safetystockthatmustbe keptby the Analysissectionbelow.
company,and q is a constantordermultiplier to provide
a bufferfor the variationin demand.(S could also be
determined bys + EOQ, whereEOQ theoptimal'econom- Analysisof thesourcesof thebullwhipeffect
ic orderquantity'to be calculatedfor giveninventory
Numeroussimulationexperiments are carriedout using
holdingand BL costs.Optimalcomputations ofs and S are
each of thethreeordering describedabove. (Some
and irrelevant
difficult to the purposeof this policies
extremely variantsof these policies and other policies like (s,Q)
- see forinstanceNahmias,S. (2009, Chapter5)).
article
have been testedas well,but we skip themdue to space
Since it is knownin generalthat EOQ is an increasing
limitations(see Gündüz, 2003)). These experiments can
function of estimateddemandEi>u usingq Eh, in lieu of
be groupedin two: policy-independent parameters the
of
EOQ is reasonable.The inventory dynamicswithLT = 3, chain and
SS = 240 and q = 3 is shownin Figure7. Once again,we supply policy-specific
parameters.
observeoscillationsand the bullwhipeffectalong the
supplychain.Note further thatthelevelof amplification
in thiscase is strongerthanthe previoustwo cases- the parameteranalysis
Policy-independent
amplitudeof oscillationsmore than doubles with each Simulationexperimentsare performedwith different
stage.As willbe analyzedlater,thisfindingis consistent settingsof demandpattern(autocorrelation
degree),LT,
with the Lee et al (1997) resultsthat 'order batching' natureof delays,and demandestimationadjustmenttime
together withdemandforecasting is one ofthemaincauses (EAT). Some importantresultscan be summarizedas

when(s,S)policyis applied(SS = 240,q = 3).


Figure7 Netinventories

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
466 oftheOperational
Journal Research Vol.62, No.3
Society

follows.Onlyone examplebehaviourgraphis showndue information delayseffectively increaseLT in receiving


to space limitation
(See Gündüz,2003). theordersplaced.
• Base retail oscillationsand bullwhipboth decrease
• If the end item demand is autocorrelated,the base significantly with an increasein the demand EAT,
(retail) oscillations and the bullwhip effect both for all three policies.(Compare Figure8 below,with
increase,especially when order-up-to-S policy or an- 4
Figure above, as an illustration.) The explanation
chor-and-adjustpolicy is used. With (s,S) policy, is that, the largerEAT, the less responsivethe model
autocorrelation does not have substantialeffecton becomesto changesin demand(or incomingorders).
inventory because significant
oscillations, 'batching'of Since uncoordinateddemand forecastingis a main
ordersin (s,S) makes it insensitive to demand auto- cause of the bullwhipeffect(Lee et al 1997), larger
correlation. But thisbatchinghas a bullwhipeffectas EAT, meaninglessresponsive (or almost'no') forecasts,
willbe seenlater. naturally lead to decreasedbullwhipeffect. Thisresultis
• The bullwhipeffectincreaseswithan increasein LT. important in the sensethatit revealsone of themajor
Thisis essentiallycausedbythefactthattheup-to-order causes of the bullwhip phenomena:uncoordinated
level S and the desiredsupply line SL, t are both demandforecasting, as willbe discussedbelow.
proportional to LT (delay).The ordersand theresulting
inventory oscillationsare, hence,amplified.The same Table 1 summarizes foreach ordering policy,theeffects
of
is not truewith(s,S) policy,wherethe orderquantity four policy-independent parameterson oscillationsand
and thefrequency are not affectedby LT. bullwhip,as discussedin thissection.
• In thebasic model,theonlydelayis thematerialdelay
on the supplyline (lead time').In reality,therecan
be otherdelayslikeinformation Policy-specificparameteranalysis
delayin placingorders.
The effectsof includingsuch additionaldelaysin the In the second set of runs,we experiment with policy-
ordering mechanisms havebeeninvestigated. In all three dependentparameterssuch as LT inflationconstantK
policies, the bullwhip effectsignificantly increases in theorder-up-to-S policy,orderquantitymultiplier q in
withinclusionof orderinformation delays(see Gündüz, (s,S) policy,and thedesiredinventory coverageconstantm
2003).Thisis consistent withourabove resultaboutthe in the anchor and adjust policy. (See the associated
bullwhipeffectsof increased LT, since additional equationsof each policy,above.)

Figure8 Netinventories policyand EAT increased


withorder-up-to-S to 10.

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
Y BariasandB GunduzDemand andsharing
forecasting to reducefluctuations
strategies andthebullwhip insupply
effect chains 467

Table 1 Summaryof policy-independent


parameteranalysis
Parameterchange: Demandpattern(fromiid Increasein lead Inclusionof order Increasein estimation
adjustment
Policy normal,to autocorr
-elated) time(from3 to 6) informationdelay time(from5 to 10)

Order-up-to-S Oscillationsandbullwhip Bullwhipeffect


increases Bullwhipeffect Oscillations
andbullwhip
effect
Policy effect
increase significantly decrease
significantly
increases
System Dynamics Oscillations
andbullwhip Bullwhipeffect
increases Bullwhipeffect Oscillations
andbullwhip
effect
Anchor-and- effect
increase significantly decrease
significantly
AdjustPolicy increases
(s,S)Policy No substantial
changein No substantial
change Bullwhip effect Oscillations
andbullwhip
effect
behaviour in behaviour significantly decrease
significantly
increases

when(s,S)policyis applied(SS = 240,q - 5).


Figure9 Netinventories

• Experiments showthatwhenLT inflation constantK (in • Whenthedesiredinventory coverageconstantm (in the


the order-up-to-S policy) is increased,the bullwhip anchor and adjust policy) is increased,the bullwhip
effectalso increases.(We are unableto showthegraphs effectand magnitudeof oscillationsboth increase
due to spacelimitations; see Gündüz,2003.)Note thatK (See Gündüz, 2003 for outputgraphs).Observethat
is a multiplier
of theexpecteddemandin theorder-up- in this policy,the desiredinventoryis obtained by
to-levelcomputation in thispolicy;so it representsthe multiplying the demand forecastswith the coverage
weightof theexpecteddemandin theorderdecisions. constantm,so thatthelatterrepresentsagaintheweight
• Whentheorderquantitymultiplier q (in (s,S) policy)is of demandexpectations in orderdecisions.
increased, thebullwhipeffect and magnitudeof oscilla-
tionsboth increase(compareFigures9 and 7). Note Table 2 summarizestheeffects oftheparameters ofeach
againthatq is a multiplier oftheexpecteddemandin the orderingpolicy,on oscillationsand bullwhip,as discussed
order-up-to-levelS formulain thispolicy;it represents in thissection.
the strengthof the demand forecastin the order All the threepolicy-specific resultspoint to a single
decisions. generalresult.The strongerthe role of isolateddemand

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
468 Journal Research
oftheOperational Vol.62,No.3
Society

ofpolicy-specific
Table2 Summary parameter
analysis
Parameter
change: inleadtimeinflation
Increase Orderquantity
multiplier
q Desiredinventory
coverage
Policy K (from1 to3)
constant (from3 to5) constantm (from3 to5)

Order-up-to-S
policy effect
Bullwhip increases
System
dynamics Bullwhip and
effect
anchor-and-adjust
policy magnitudeofoscillations
increase
(s,S)policy Bullwhipeffect
and
magnitudeofoscillations
increase

under(s,S)policy,without
Figure10 Netinventories usinganyforecasting.

forecasts in theorderdecisions,thestronger is thebullwhip Figure7 withthosein Figures4 and 6. Buttheothermajor


effect (and in somecases thebase retailoscillations).This cause of thebullwhip,ie demand/order is also
forecasting,
resultis consistentwithForrester(1961), Sterman(2000), presentin these(s,S) policyruns.In orderto isolateand
Ya§arcanand Barlas(2005),Barlasand Özevin(2004),Lee focuson the effectof orderbatchingonly,we test(s,S)
et al (1997),and Chen et al (1998). This finding,together policywithfixedlevelsof s and S thatare not updated
withtheearlierresultabout therole of orderbatchingin by any demand/order forecasts.Results in this case
amplifying the bullwhipeffect,leads to two important (Figure 10, compared to Figure 7) reveal that when
policy-oriented rootsof thebullwhip. demandforecastsare not used, undertheeffectof order
batchingonly,thebullwhipdoes notpropagatethrough the
entirechain,it occursfromthe retailerto thewholesaler
The role of orderbatching
only.The amplification fromtheretailerto thewholesaler
Amongthepoliciestested,ordersare batchedin the(s,S) is a resultof batched orders,so the wholesalerfaces
policy,sincetheyare notplacedin each period.We noted batchedorderswhilethe retailerfacesun-batchedretail
earlierthattheeffectof thisbatchingis an increasein the demand. There is no amplificationat all from the
bullwhipeffect,comparingthe largeramplifications in wholesalerto theproducer,because both of theseagents

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
- Demand
Y BarlasandB Gunduz andsharing
forecasting to reducefluctuations
strategies andthebullwhip insupply
effect chains 469

face the same type of batched orders,with the same Improvedpolicy:demandand forecastinformation
ordering rules.The samewouldbe truefromanystagen to sharing
stagew+1, n>'). Thus, we concludethat the batching
of ordersby itselfis notsufficient forthe bullwhipeffect
Thereare severalstrategies suggestedin the literature to
tacklethebullwhipeffect, as summarized in theintroduc-
to propagatein thesupplychain;orderpolicyparameters
tion. One such strategy, also impliedby our results,is
mustbe updatedby demandforecasts forthebullwhipto
that sharingof demandand/orforecastinformation between
propagate. But we also conclude batchingcan further
if the agentsin thesupplychain.Thereis a richliterature arguing
the
amplify bullwhipeffect, latteralreadyexistsas a fordemand/forecast information
resultof demand/order forecasts.Finally,note that if sharing,some of which
was mentioned above in theIntroduction section.In order
agentsused higherand higherdegreesof batching(ie less to explorethe effectsof this strategyon the behaviour
frequentorders)as we upstream,then order batching of the inventories, we modifythe supplychainmodel to
couldbyitselfresultin bullwhipeffects thatwouldpropa-
even without incorporateend-itemdemand sharing.Each stage uses
gate through the entire chain, any demand end-itemdemand informationto forecastthe future
forecasting.Such detailed analysis of the specificsof
is the demand,ratherthanusingordersofitslowerstage.Hence,
different orderingpolicies beyond scope of this all stages use demand forecastsobtaineddirectlyfrom
article.For instance,Potterand Disney(2006)showthatin
end-itemdemandin theirorderingdecisions.In the base
(S, nQ) policy,thebullwhipeffectis reducedif thebatch case reportedhere,sinceall agentsin the model use the
sizeQ is a multipleofaveragedemand,and in-between the
same forecastingmechanismwith same parameters,
minima,thebullwhipeffect risesand fallsin a waveform,
end-item demandsharingis equivalentto demandforecasts
reachinga peak at thehalfwaypoint. Given that the end itemdemandis shared,all
sharing.
stages effectively produce and use the same end item
The role of demandi orderforecasting demand forecasts.(If agents used different forecasting
methods, thendemandsharingand forecastsharingwould
Analysisof the simulationresultsof no-demand-sharing be two different
cases revealsthata primarycause of the bullwhipeffect strategies.)
The resultinginventorybehavioursfor order-up-to-S
is the isolatedsequentialdemandforecasting performed SD and (s,S) policywhendemandis shared
at each stageof thesupplychain,makinguse of previous policy, policy,
are shown in Figures 11-13 respectively. Demand and
stage'sorders.This resultis also supportedby Lee et al forecast sharing eliminates uncoordinatedsequential
(1997),Chen et al (1998),Sterman(2000,Chapter17 and mechanisms of thesupplychainso thata stage
forecasting
18), and Forrester(1961, Chapter12). As we have seen no bases its orders on its forecastsof the lower
longer
above, the weight of demand forecastsin ordering
decisionsdetermines thedegreeofthebullwhipexperienced stage's
orders. Instead, each stage directlyuses end-item
forecasts.Thus, the bullwhip effectalongthesupplychain
by the chain. All orderingpolicies that use demand is
forecasts in ordering demandforecasts
reduced.
significantly (CompareFigure11 to Figure4;
equationsmultiply
Figure 12 to Figure6; and Figure 13 to Figure7.) Our
by a constantin order to obtain some 'target'order resultsare consistentwith
level.This constantis K fororder-up- to-S policy,m for experimentaland empirical
evidenceon the role of demandinformation sharingon
anchor-and-adjust policy, and q for (s,S) policy (see
stabilizingsupply chain inventories (for instance,Chen
equationsabove). Simulationrunsshowthatthehigheris et al, 1998;Chengand Wu, 2005; Croson and Donohue,
themultiplier constant,thegreateris themagnitude of the
effect 9 for for the 2005).
bullwhip (see Figure example, (s,S)
policycase).
Experiments withdifferent EAT valuesalso revealthat
the bullwhipeffectdecreaseswith an increasein EAT Order-up-to-S policy
(Figure8). Increasein EAT meansthatdemandforecasts Anchorand
arelessresponsive to changesin demand.A veryhighEAT adjustpolicy
effectively means no forecastupdating,yielding'almost
constant'demand forecasts.At the extreme,constant (s,S) policy
('no') demandforecasting resultsin zero bullwhipeffect Anotherstrategy suggestedagainstbullwhipis theechelon
(Figure10 forexample).Note thatthislastextremeresult inventory policy,whereeach agentplacesordersbased on
is of theoretical
value,butwouldnot be implementable in echelon IP ratherthat its local position.The echelon
real world.Inventory management withoutany demand inventory of a stage is definedas IP of the subsystem
forecasting wouldobviouslycausemajorproblemsin terms consisting of thestageitselfand all itsdownstream stages
ofshortages and overstocking, so eliminatingthebullwhip (see Silveretal, 1998).Whenthemodelis runwithEchelon
effectwould not be of muchpracticalvalue withsuch a policies,we obtaina further decreasein thebullwhipeffect
strategy. becausethesepoliciesremovetheorderpropagationdelay

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
470 Journal
oftheOperational
Research Vol.62, No.3
Society

Figure11 Netinventories
whenorderup-to-Spolicyis appliedanddemandis shared.

whenanchorandadjustpolicyis appliedanddemandis shared.


Figure12 Netinventories

fromthesupplychain.(We are unableto provideoutput Table 3 summarizesthe consequencesof the demand


graphsand further discussionof echelonpoliciesdue to forecasting/sharingand the Echelon inventorypolicies
space limitations
(see Gündüz,2003)). vis-a-visthebullwhipeffect.

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
Y BariasandB Gunduz-
Demand andsharing
forecasting to reducefluctuations
strategies andthebullwhip insupply
effect chains 47 1

when(s,S) policyis appliedand demandis shared.


Figure13 Net inventories

Table 3 Summaryresultsof information-sharing


policies
Scenario: No demandforecastis Demandandforecastsharingis policyis used
Echeloninventory
employed employed

Order-up-to-S
policy No bullwhip
effect effect
Bullwhip significantly Further in bullwhip
decrease
decreases effect
System dynamics No bullwhip
effect effect
Bullwhip significantly Further in bullwhip
decrease
anchor-and-adjust decreases effect
policy
(s,S)policy No bullwhip
effect effect
Bullwhip significantly Further in bullwhip
decrease
decreases effect

There are several well-knownbusiness practicesto structure.Disney et al (2004) analyze the impact of
implement demandinformation-sharing and/orEchelon- information and communicationtechnologies(ICT) on
inventorypolicies: Collaborative Planning, Vendor thebullwhipand supplychainperformance. Theyconclude
Managed Inventories (VMI), Continuous Replenishment that althoughthere are benefitsof ICT on bullwhip
Programs(CRP), information systems likeElectronicData and supplychain performance, such policies should be
Interchange (EDI), Pointof Sale (POS) applications,and implemented carefully because of the added complexities
more recentdevelopments like Web-basedTransactions, to thedecisionmakingprocess.Holwegetal (2005)classify
and Radio FrequencyIdentification (RFID) tags in lieu supplychain collaborationinitiatives based on inventory
of bar codes (see Nahmias,2009 and Lee et al, 1997 and and
replenishment forecasting collaboration dimensions.
1998). Disney SM and Towill DR (2003) analyze the They claim that collaborationin inventoryreplenishment
effectsof VMI structures on two particularsourcesof alone (eg VMI) or forecastsharingalone is not sufficient
the bullwhipeffect(demandsignalprocessingand order to achievesignificant improvement in thebullwhipeffect.
lead times).Theydemonstrate by simulationexperiments Sharedinformation shouldbe usedin supplier'sforecasting
that the bullwhipeffectis reducedsignificantly if VMI and inventory controlprocessesin orderto gainsubstantial
structures are comparedto the traditionalserial chain improvements in thebullwhip.

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
472 Journal Research
oftheOperational Vol.62,No.3
Society

Conclusionsand futurework like EDI, POS applications,Web-based Transaction


Systems,and RFID tags in lieu of bar codes. To avoid
Threetypicalorderingpoliciesare consideredand mod- excessivebatching,a contributor CRP can be
to bullwhip,
elled in the context of a supply chain. Numerous used.To shortentheLT, anotherbullwhipamplifier,Quick
simulationexperiments are carriedout usingeach of the Response(QR) systems can be implemented.
three policies. These experimentsanalyze two groups Thereare two othermajorcauses of thebullwhipeffect
of factors:policy-independent parametersof the supply knownin the literature:shortage(rationing)gamingand
chainand policy-specific parameters foreach policy. price variations.
These two factors,necessitatingmore
The mostgeneralconclusionof theexperiments is that detailedand complexmodelling,are beyondthescope of
the bullwhipeffect(amplification of orders along the thisstudyand constitutefurther researchareas. Finally,
supplychain) resultsin all cases and policies,as long as informationsharing on supply networkstructuresis
each stageutilizeslocal uncoordinated forecastsbased on anotherpotentialresearchtopic. Advanced forecasting
incoming orders (or demand). So, uncoordinated local modelsinvolving methodsand moresophis-
extrapolative
demandforecasting is confirmed to be as a maincause of ticatedorderingpoliciesand can be testedin such more
thebullwhipeffect. An extensionof thisresultis thatthe realisticand complexsettings.
level of 'responsiveness'of forecaststo the demand
influences the magnitudeof the bullwhipeffectexperi- Acknowledgements - Supportedby Bogazici University Research
enced.Forecaststhatare highlyresponsive to thechanges Fundno. 02R102.
in the demand increasethe bullwhipeffect,while less
responsiveforecastsdecreaseit. The weightof demand References
forecastsin the orderingequation is anotherimportant
factorthatdetermines thebullwhipeffect. If theweightof Barias Y (2002). Systemdynamics:Systemicfeedbackmodeling
forpolicyanalysis.In: Knowledge forSustainable
Development -
demandforecastsin orderingequationsis high,thenthe
An Insightinto the Encyclopediaof Life SupportSystems.
magnitudeof bullwhipis also high.If forecastsare not UNESCO-EOLSS Publishers:Paris, France, Oxford,UK,
used in orderingequationat all, thenthe bullwhipeffect pp 1131-1175.
does not exist.Thus, if the demand patternis known BarlasY and ÖzevinMG (2004). Analysisof stockmanagement
gamingexperiments and alternative formulations.
ordering Syst
(and/oritchangesveryslowly),thebullwhipeffect maybe
Res BehavSci 21: 439^70.
avoided by not using the demandforecastsin ordering
Chen F, DreznerZ, Ryan JK and Simchi-Levi D (1998). The
equationsor by usingveryslow-response forecasts. bullwhip effect:
Managerialinsights on theimpactofforecasting
Two otherfactorswerepartiallyanalyzedby simulation and information on variabilityin a supplychain.In: TayurS,
experiments: supply line LT and batchingof orders. GaneshanR and MagazineM (eds). Quantitative Modelsfor
show that LT itself is not sufficient to Supply Chain Management.Kluwer Academic Publishers:
Experiments by
createthebullwhipeffect.But giventhatthereis already Boston,pp 417^40.
Chen F, Ryan JK and Simchi-LeviD (2000). The impactof
a bullwhipeffectcaused by local demandforecasting (or exponential smoothing forecasts on thebullwhipeffect.
NavRes
orderbatchingor shortagegaming),LT does amplifythe Log 47: 269-286.
bullwhipsubstantially.Increased level of batchingof ChengTCE andWu YN (2005).Theimpactofinformation sharing
ordersand increasedLT, both cause in a two-levelsupplychainwithmultiple J Opl Res Soc
retailers.
amplifiedbullwhip
56: 1159-1165.
effects.Batchingof ordersby itselfis sufficient to cause
CrosonR and Donohue K (2005). Upstreamversusdownstream
an orderamplification fromtheretailerto thenextstage, information and itsimpacton thebullwhipeffect. SystDynam
thewholesaler.But thisamplification is limited,as it does Rev21: 249-260.
notpropagateto theotherstages,ifall agentsuse thesame Datta S, GrangerCWJ,BarariM and GibbsT (2007). Manage-
degreeof batching.For the bullwhipcaused by order mentof supplychain:An alternative modellingtechniquefor
forecasting.J Opl Res Soc 58: 1459-1469.
batchingto propagatethroughtheentirechain,agentsup DejonckñeereJ, Disney aM, LambrechtMK and lowill JJK
streammustuse increasing degreesof orderbatching. (2002). Transferfunctionanalysis of forecastinginduced
Lastly,we testeddemand and forecastinformation- bullwhipin supplychains.IntJ ProdEcon78: 133-144.
sharingstrategies againstthebullwhipeffect, as suggested DejonckheereJ, Disney SM, LambrechtMR and Towill DR
These strategies reducethe (2004).The impactof information enrichmenton thebullwhip
by the literature. significantly effectin supplychains:A controlengineering EurJ
in demand perspective.
bullwhipeffect all orderingpolicies.However, Opl Res 153: 727-750.
and forecastsharingcannot completelyeliminatethe DisneySM and TowillDR (2003).The effectof vendormanaged
bullwhip;it can reduceit. Bullwhipeffectwill exist to inventory (VMI) dynamicson the BullwhipEffectin supply
some extentin the supplychain as long as the ordering chains.IntJ ProdEcon85: 199-215.
policiesuse isolateddemandforecasts (or thereis shortage Disney SM, NairnMM and PotterA (2004).Assessingtheimpact
of e-businesson supplychaindynamics.Int J Prod Econ 89:
gaming,or price variations). The information-sharing 109-118.
strategyimpliesmanagerialpracticeslike Collaborative ForresterJW (1961). Industrial Dynamics.PegasusCommunica-
Planningand VMI. It also necessitates information systems tions:Waltham,Mass.

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions
Y BariasandB Gunduz-
Demand andsharing
forecasting strategies andthebullwhip
to reducefluctuations insupply
effect chains 473

GallegoG and ÖzerÖ (2001).Optimaluse ofdemandinformation NahmiasS (2009). Production and OperationsAnalysis.McGraw-
in supplychainmanagement. In: Song J and Yao DD (eds). Hill:New York.
SupplyChainStructures: Coordination, and Optimi-
Information PotterA and Disney SM (2006). Bullwhipand batching:An
zation.KluwerAcademicPublishers: Boston,pp 127-168. exploration.IntJ ProdEcon 104:408^418.
GavirneniS, KapuscinskiR and Tayur S (1999). Value of SilverEA, PykeDF and PetersonR (1998).InventoryManagement
information supplychains.ManageSci 45: 16-24.
incapacitated andProduction PlanningandScheduling.3rdedn,JohnWiley&
Gündüz AB (2003). Strategiesto reducefluctuations in supply Sons:New York.
chains:A dynamic modelingapproach.MS Thesis,Department of Sterman JD (1989).Modelingmanagerialbehavior:Misperceptions
IndustrialEngineering,BogaziciUniversity. of feedbackin a dynamicdecisionexperiment.ManageSci 35:
HolwegM, DisneySM, Holmstrom Jand SmarosJ(2005).Supply 321-339.
chaincollaboration:Makingsenseof the strategy continuum. StermanJD (2000).BusinessDynamics. McGraw-Hill:Boston.
EurManageJ 23: 170-181. Xu K, Dong Y and EversPT (2001).Towardsbettercoordination
JeongS and Maday CJ (1996). Dynamicinformation controlfor of thesupplychain.TransportRes E 37: 35-54.
multi-echelon systems
production-distribution withconstrained Ya§ arcan H and Barlas Y (2005). A generalizedstockcontrol
production capacity.SystDynamRev12: 331-343. formulation forstockmanagement problemsinvolvingcompo-
Lee HL, PadmanabhanP and Whang S (1997). Information sitedelaysand secondarystocks.SystDynamRev21: 33-68.
distortionina supplychain:The bullwhip ManageSci 43:
effect.
546-558.
Lee HL andWhangS (1998).Information ina supplychain.
sharing ReceivedApril2009;
Research Paper,Stanford GraduateSchoolofBusiness.
University acceptedOctober2010

This content downloaded from 129.16.69.49 on Fri, 08 Jan 2016 15:25:08 UTC
All use subject to JSTOR Terms and Conditions

Das könnte Ihnen auch gefallen