ARTICLE IN PRESS

Int. J. Production Economics 118 (2009) 311322

Contents lists available at ScienceDirect

Int. J. Production Economics

journal homepage: www.elsevier.com/locate/ijpe

The bullwhip effect in supply chainsAn overestimated problem?

Eric Sucky
Chair of Operations Management and Business Logistics, University of Bamberg, Feldkirchenstr. 21, 96052 Bamberg, Germany

a r t i c l e in fo abstract

Available online 23 August 2008 A phenomenon that is now well known as the bullwhip effect suggests that the
Keywords: variability of orders increases as they move up the supply chain from retailers to
Bullwhip effect wholesalers to manufacturers to suppliers. In this paper, we will focus mainly on
Risk pooling measuring the bullwhip effect. Existing approaches that aim at quantifying the bullwhip
Supply network effect neglect the network structure of supply chains. By only assuming a simple two-
Order-up-to policy stage supply chain consisting of a single retailer and a single manufacturer, some of the
Square root law relevant risk pooling effects associated with the network structure of supply chains are
disregarded. Risk pooling effects arise when the orders, which a retailer receives from its
customers, are statistically correlated with a coefcient of correlation less than one.
When analyzing the bullwhip effect in supply chains, however, the inuence of risk
pooling has to be considered. The fact that these inuences have not yet been analyzed
motivates the research presented in this paper. We will show that the bullwhip effect is
overestimated if just a simple supply chain is assumed and risk pooling effects are
present.
& 2008 Elsevier B.V. All rights reserved.

1. Introduction (Lee et al., 1997b). In another paper, the same authors, Lee
et al. (1997a), observe the bullwhip effect in a soup supply
A phenomenon that is now well known as the bullwhip chain as well as in the supply chain for printers of
effect suggests that the variability of orders increases as HewlettPackard (HP). Barilla also nds that phenomenon
we move upstream in the supply chain from retail to in the supply chain for pasta (Hammond, 1994). Further-
manufacturing. In a supply chain, although consumer more, Terwiesch et al. (2005) have found that the
sales do not seem to vary much, there is pronounced semiconductor equipment industry is more volatile than
variability in the retailers orders to the wholesaler. the personal computer industry, and Anderson et al.
Furthermore, the wholesalers order quantities to the (2000) assign the volatility in the machine tool industry to
manufacturer as well as the manufacturers orders to the the bullwhip effect. Additionally, the bullwhip effect has
supplier vary even more in time. For a detailed elaboration been experienced by many subjects playing The Beer Game
of the bullwhip effect, see Kahn (1987), Lee et al. (Sterman, 1989).
(1997a, b), and Metters (1997). Regarding the large number of studies, which observed
Numerous studies nd the bullwhip effect in some an increase in demand variability as one moves up a
industries and in numerous examples from individual supply chain, Lee et al. (2004) conclude that nowadays
products and companies. In the supply chain for diapers, the bullwhip effect is a standard industry term and
Procter and Gamble (P&G) noticed that the volatility of the reference to it in industry publications has become
diaper orders issued by the distributors was quite high commonplace (p. 1891). In a seminal paper, these same
even though end consumer demand was reasonably stable authors identify four major causes of the bullwhip
effect(i) the updating of demand forecasts, (ii) order
batching, (iii) price uctuation, and (iv) rationing
E-mail address: [email protected] and shortage gamingand suggest several managerial

0925-5273/$ - see front matter & 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.ijpe.2008.08.035
 ARTICLE IN PRESS
312 E. Sucky / Int. J. Production Economics 118 (2009) 311322

practices to mitigate its consequences. In addition, We extend the analysis of Chen et al. (1999, 2000) to a
Dejonckheere et al. (2003) nd that an important factor supply chain with a network structure in which risk
to the bullwhip effect is the replenishment rule used by pooling can reduce the bullwhip effect on every individual
the supply chain members. The authors conclude that stage. We rst describe the supply chain setting, the
whatever forecasting method is used (e.g. exponential forecasting technique, and the inventory policy used by
smoothing or moving averages), order-up-to systems will the individual actors. To measure the bullwhip effect, the
always result in the bullwhip effect. variances of the orders placed by the wholesalers to the
However, in a recent study of US industry level data, manufacturers relative to the variance of the demand
Cachon et al. (2005) nd that demand volatility does not faced by the wholesalers will be determined. We will
increase as one moves up the supply chain. In contrast to show that the bullwhip effect may be overestimated if just
the natural consequences of the bullwhip effect, the a simple supply chain is assumed and risk pooling effects
authors observe thatin generalmanufacturers do not are present. Therefore, in a supply network, using a simple
have substantially greater demand volatility than retailers forecasting method (e.g. moving averages), order-up-to
and may have even lower demand volatility. These results systems will not always result in the bullwhip effect. The
are explained mainly by production smoothing: predict- analytical results will be illustrated and afrmed by a
able seasonality in combination with increasing marginal simulation study.
costs provides a strong motivation to smooth production
relative to demand. Therefore, the majority of retail and
manufacturing industries smooth their production rela- 2. Related literature
tive to their demand, i.e., they impose less volatility on
their suppliers than they face from their customers Since the rst analysis of this phenomenon by
(Cachon et al., 2005, pp. 1819). Cachon et al. (2005) Forrester (1958, 1961), the bullwhip effect has been
conclude that the bullwhip effect is not widespread in the addressed in a large number of publications. Recent
US economy and, moreover, the bullwhip effect is not research on the bullwhip effect can be divided into six
commonplace. general categories: (i) papers aiming at a quantication of
In this paper, we analyze another strong force that the bullwhip effect (e.g. Carlsson and Fuller, 2000; Chen et
mitigates the bullwhip effect. We focus on analyzing and al., 2000; Dejonckheere et al., 2003; Kahn, 1987; Lee et al.,
measuring the bullwhip effect analytically. Because 1997a, b; Metters, 1997; Zhou and Disney, 2006), (ii)
supply chains are more like supply networks, our works focusing on analyzing and identifying the causes of
analysis accounts for supply chains that possess a network the bullwhip effect (e.g. Geary et al., 2006; Lee et al.,
structure. In practice, supply chains can be considered as 1997a, b; Metters, 1997; Nienhaus et al., 2006), (iii)
networks of geographically dispersed facilitieswhere studies observing the bullwhip effect in some industries
raw materials, intermediate and nished products are or in numerous examples from individual products and
produced, tested, modied, and storedand the trans- companies (e.g. Cachon et al., 2005; Lee et al., 1997a), (iv)
portation links that connect the facilities. The different papers addressing methods for reducing the bullwhip
operations (e.g. raw materials procurement, nished effect (e.g. Carlsson and Fuller, 2001; Chen et al., 1999;
goods manufacturing, and distribution) are performed Dejonckheere et al., 2003; Disney and Towill, 2003; Ingalls
on different stages of the supply chain. The term supply et al., 2005; Mason-Jones and Towill, 2000; Moyaux et al.,
chain implies that only one player is involved at each 2007), (v) works focusing on simulating the system
stage of the supply chain. In reality, however, a manu- behavior (e.g. Disney and Towill, 2003; Ingalls et al.,
facturer supplies several wholesalers and may receive 2005; Makajic-Nikolic et al., 2004; Nienhaus et al., 2006),
material from several suppliers. Therefore, in the follow- and (vi) papers focusing on experimental validation of the
ing, we use the term supply chain if only one player is bullwhip effect (e.g. Moyaux et al., 2003).
involved at each stage, i.e. if the supply chain has a linear A great part of previous research has focused on
structure. If two or more players are involved at one stage, demonstrating the existence and identifying the possible
we will consequently use the term supply network. causes of the bullwhip effect (category (ii) of the relevant
Existing approaches that aim at quantifying the literature). Particularly, Lee et al. (1997a, b) identify four
bullwhip effect neglect the network structure of supply major causes of the considered phenomenon: (a) the
chains. By assuming only a three-stage supply chain updating of demand forecasts, (b) order batching, (c) price
consisting of a single customer, a single retailer, and a uctuation, and (d) rationing and shortage gaming. The
single manufacturer, some relevant risk pooling effects rst cause occurs when the parties involved in the supply
associated with the network structure of supply chains are chain base their forecasts on the historical demand
disregarded. Risk pooling effects arise for example when behavior of their immediate customers. Every supply
the orders a retailer receives from its customers are chain member then adjusts to uctuations of their order
statistically correlated with a coefcient of correlation less entry. Moreover, if every member reacts to uctuations
than one. Note that the risk pooling effect is a special case with smoothing techniques, the uctuations will amplify
of the well-known portfolio effect (Ronen, 1990). When throughout the supply chain. The effect of order batch-
analyzing the bullwhip effect in supply chains, however, ingwhich is a rational order policy if the costs for
the inuence of risk pooling cannot be neglected. The fact frequent order processing are highis an amplication of
that these inuences have not been analyzed yet moti- the order variability; the connection between the order
vates the research presented in this paper. policy and the actual demand patterns of the customers is
 ARTICLE IN PRESS
E. Sucky / Int. J. Production Economics 118 (2009) 311322 313

then imparted. In case of price uctuations customers are information flow

driven to buy in larger quantities by attractive offers,
which may also include quantity discounts or price
discounts. The resulting buying patterns will not reect
consumption patterns anymore, i.e. customers buy in Manufacturer
Manufacturer Wholesaler
Wholesaler Retailer
quantities, which do not reect their needs. Finally, the
rationing and shortage game occurs when demand
material flow
exceeds supply. The customers may start to exaggerate
in comparison to their actual needs when there is a fear Fig. 1. The three-stage supply chain.
that supply will not cover demand (Carlsson and Fuller,
2000). These four causes are interdependent; the causes with periodic review, where Dt is the stochastic and
may interact and act in concert. However, the updating of stationary demand of the retailer in any period t. By
demand forecasts appears to be the major source of the stationary, we imply that the retailers demand in
bullwhip effect. different time periods uctuates randomly around a
The bullwhip effect has a number of negative effects in constant mean level. We assume that the retailers
real supply chains, which can cause signicant inefcien- demands are independent over time and identically
cies. The bullwhip effect typically leads to excessive distributed random variables. Assuming that a review is
inventory investments throughout the supply chain as made at the start of each period tA{1,2, y, T}, under an
the parties involved need to protect themselves against order-up-to policy the wholesaler places an order qt to the
demand variations. Therefore, in another class of papers, manufacturer considering the target inventory level yt for
methods reducing the bullwhip effect are proposed. While period t. Afterwards, the wholesaler lls the retailers
Lee et al. (1997b) suggest several managerial practices demand Dt for period t from on-hand inventory. Note that
to reduce the bullwhip effect (e.g. the centralization any unlled demandsthe shortagesare backlogged
of demand information), other papers specialized on (except at the end of period T, when they are lost). In
using forecasting methods to reduce its consequences order to focus on the inuence of risk pooling on the
(e.g. Carlsson and Fuller, 2000; Dejonckheere et al., 2003). bullwhip effect, we set, in contrast to Chen et al. (2000),
Carlsson and Fuller (2000) suggest a fuzzy approach to the lead time to zero, which means that any order placed
estimate future demand. at the start of period t is also available at the beginning of
The category (i) of papers analyzing the bullwhip effect period t. This is how we can quantify the bullwhip effect
is focused on quantifying the increase in variability at without the inuence of lead times.
each stage of the supply chain (e.g. Chen et al., 2000; Lee In order to determine the orders qt, similar to Chen et
et al., 1997b; Metters, 1997). However, these approaches al. (2000), we assume that the wholesaler follows an
neglect the more complex network structure of real order-up-to inventory policy. The goal of this ordering
supply chains. By assuming only a simple two-stage policy is to bring the actual inventory towards the desired
supply chain consisting of a single retailer and a single inventory yt (Johnson and Montgomery, 1974). The order
manufacturer, some relevant risk pooling effects associated quantity qt, the wholesaler places to the manufacturer at
with the network structure of supply chainstherefore the start of period t, is given by
often called supply networksare disregarded. Risk
pooling effects arise, for example, when the orders a qt yt  yt1 Dt1 , (1)
retailer receives from its customers are statistically
where yt is the desired order-up-to level (target inven-
correlated with a coefcient of correlation less than one.
tory), yt1 the order-up-to level at the end of period t1,
When analyzing the bullwhip effect in supply networks,
and Dt1 the perceived demand. In case of qto0 we
the inuence of risk pooling has to be considered. In the
assume that this excess inventory is returned to the
following, we will show that the bullwhip effect is
manufacturer without cost, i.e. we allow costless returns
overvalued if just a simple supply chain is assumed and
(Kahn, 1987; Lee et al., 1997b). According to Chen et al.
risk pooling effects in supply networks are present.
(2000), we assume that the wholesaler follows a simple
order-up-to inventory policy in which the order-up-to level
3. The bullwhip effect in supply chains (target stock level) in period t is estimated from the
observed demand as
The analysis of the bullwhip effect will be based on the p
approaches of Chen et al. (1999, 2000), Metters (1997), Lee yt EDt z VarDt , (2)
et al. (1997b), and Kahn (1987), which consider a two- p
where E(Dt) is an estimate of the mean demand, VarDt
stage supply chain. In this paper, we will analyze a three-
an estimate of the standard deviation of the retailers
stage supply chain consistingat rstof a single retailer,
demand, and zX0 is the safety factor chosen to meet a
a single wholesaler, and a single manufacturer (see Fig. 1).
desired service level (Silver et al., 1998). Note that z is a
managerial determined factor that indicates the number
3.1. The inventory policy of estimated standard deviations of demand to be kept as
safety stock (Zinn et al., 1989). For z 0 the decision
The considered inventory context is the following. We maker is risk neutral, and for z40 the decision maker is
assume an inventory system managed by the wholesaler risk averse.
 ARTICLE IN PRESS
314 E. Sucky / Int. J. Production Economics 118 (2009) 311322

3.2. The forecasting technique 3.3. Quantifying the bullwhip effect in supply chains

To estimate E(Dt) and Var(Dt)following Chen et al. To quantify the bullwhip effect Chen et al. (2000)
(2000)the wholesaler uses the simple N-period moving suggest using Var(q)/Var(D), where Var(D) denotes the
average. The estimated mean of the retailers demand in variance of the retailers demand and Var(q) refers to
period t is given by the variance of orders placed by the wholesaler. In the
  X   example presented above, regarding the order quantities
1 t1 1
EDt Di Dt1 Dt2    DtN , q3, y, q20 and the demands D3, y, D20, the resulting
N itN N
bullwhip effect (for the regarded periods) is given by
i.e. the mean of the N most recent observations is used as the Var(q)/Var(D) 2266.63/265.27 8.54 (see Fig. 2).
forecast for the next period. We will use the notation MA(N) In order to analytically quantify the increase in
for N-period moving averages. The estimated variance of the variability from the wholesaler to the manufacturer, i.e.
retailers demand in t is given by to quantify the bullwhip effect, we rst determine the
  X   variance of the orders placed by the wholesaler to the
1 t1 1
VarDt Di  EDt 2 Dt1  EDt 2 manufacturer. Using the forecast method presented in
N itN N
Section 3.2, the order quantities qt are given by (3). Based
   DtN  EDt 2 . on (3) the estimated variance of the wholesalers order
Note that the estimated values of E(Dt) and Var(Dt) may quantity in t is given by
change every period and, hence, the wholesalers order-up-to    
level changes also every period (Simchi-Levi et al., 2000). 1 2 1 2
Varqt 1 VarDt1  VarDt1N
Given the estimates of the mean demand E(Dt) and the N N
p p p
standard deviation VarDt we can write the order quantity z2 Var VarDt  VarDt1
qt as  
p p 1 1
qt EDt z VarDt  EDt1  z VarDt1 2 1 CovDt1 ; Dt1N
! N N
  X   
1 t1 Xt2
2 p
Dt1 Di  Di 2z 1 Cov Dt1 ; VarDt . (4)
N itN it1N N
p p
z VarDt  VarDt1 Dt1
  p Retailer's demand
1 140
Dt1  Dt1N z VarDt Wholesaler's orders
N
p 120
 VarDt1 Dt1
    100
1 1
1 Dt1  80
N N
p p
Quantity

60
Dt1N z VarDt  VarDt1 . (3)
40
The following illustrative example shows the determina- 20
tion of the orders placed by the wholesaler depending on
the periodical demands of the retailer. In this example, we 0
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
use a 2-period moving average MA(2) and a safety factor -20
(z value) of z 2.33 (representing the desired service Time
-40
level). The calculation was performed on a Microsoft
Excels spreadsheet (Table 1). Fig. 2. Illustration of the bullwhip effect.

Table 1
Determination of wholesalers order quantities

Period 0 1 2 3 4 5 6 7 8 9 10

Dt 50 81 39 33 49 61 50 83 44 57 46
E(Dt) 65.5 60 36 41 55 55.5 66.5 63.5 50.5
Var(Dt) 240.25 441 9 64 36 30.25 272.25 380.25 42.25
yt 101.62 108.93 42.99 59.64 68.98 68.32 104.95 108.94 65.65
qt 46.32 32.94 65.65 70.34 49.34 119.63 47.99 13.71

Period 11 12 13 14 15 16 17 18 19 20

Dt 79 41 31 52 83 42 31 72 62 55
E(Dt) 51.5 62.5 60 36 41.5 67.5 62.5 36.5 51.5 67
Var(Dt) 30.25 272.25 361 25 110.25 240.25 420.25 30.25 420.25 25
yt 64.32 100.95 104.27 47.65 65.97 103.62 110.27 49.32 99.27 78.65
qt 44.67 115.63 44.33 25.62 70.32 120.65 48.65 -29.95 121.95 41.39
 ARTICLE IN PRESS
E. Sucky / Int. J. Production Economics 118 (2009) 311322 315

Note that the estimated values of Var(qt) and Var(Dt) tion facilities, warehouses, and transportation links con-
may change every period. However, the bullwhip effect is necting the aforementioned locations. The supply chain
measured by the quotient Var(q)/Var(D), i.e. in order to can be subdivided into different stages where different
quantify the bullwhip effect we have to determine the operations (e.g. raw materials procurement, nished
variances of the wholesalers order quantity and the goods manufacturing, and distribution) are performed.
retailers demand for the complete planning period. In general, the operations performed at each stage of the
Considering the planning period [1,T] with the order supply chain are distributed among several geographically
quantities q1, y, qT and the realized demands D1, y, DT, dispersed facilities owned by different companies. The
the sample variances and are given by Varq number of stages, the number of facilities at each stage,
P P
1=T Tt1 qt  Eq2 , with Eq 1=T Tt1 qt , and and the number of links between the locations determine
PT P the network structure of the supply chain and conse-
VarD 1=T t1 Dt  ED2 , with ED 1=T Tt1 Dt .
quently also the material ow from the raw materials
Regarding the specic structure of the estimated
stage to the nal customer stage.
variance of the wholesalers order quantity at each period
Extending the approach presented above, we assume a
t 2 f1; 2; :::; Tgas shown in (4)a lower bound on the
three-stage supply network (see Fig. 3) consisting of two
variance for the complete planning period can be
retailers, a single wholesaler, and a single manufacturer.
calculated (see Chen et al. (2000), p. 438). Chen et al.
We refer to the rst retailer as party (R1) and use subscript
(2000) show that
 R1 to designate his set of parameters. Similarly, we refer
p
Cov Dti ; VarDt 0 8 i 1; . . . ; N. to the second retailer in our supply network as party (R2),
and, hence generally use subscript R2 to designate his
Assuming that the retailers demands are stochastically set of parameters. In each period t, the wholesaler places
independent, i.e. the demands are independent and an order qt to the manufacturer. Afterwards the whole-
identically distributed, follows Cov(Dt1,Dt1N) 0. In saler lls the demands for that period, denoted by DR1,t for
order to calculate a lower bound on the wholesalers retailer (R1) and DR2,t for retailer (R2). The overall demand
PJ
orders variance, we furthermore assume z 0. So, we are for period t is given by j 1DRj,t, with J 2. As in the
able to determine a lower bound on the variance of the previous section, we assume that the wholesaler follows
wholesalers order quantity for the whole planning period a simple order-up-to inventory policy (see Eqs. (1), (2),
as and (3)).
   
1 2 1 2
Varq 1 VarD  VarD
N N 4.1. The risk pooling effect in supply networks
 
2 2
1 2 VarD. (5)
N N To motivate our further analysis, we rst show a simple
Furthermore, we are able to determine a lower bound statistical effect. If the wholesaler calculates the order-up-
on the increase in variability from the wholesaler to the to level for the retailers demands in an isolated manner
p
manufacturer, i.e. the bullwhip effect, as (see Chen et al. it followsp
yR1;t EDR1;t z VarDR1;t and yR2;t
(2000), pp. 438439) EDR2;t z VarDR2;t . In this case, the overall order-up-
  to level in period t is given by
Varq 2 2
X 1 2 . (6)
VarD N N
yt yR1;t yR2;t EDR1;t EDR2;t
The lower bound (6) is tight for z 0, i.e. if the q q
wholesalers target stock level is solely depending on the z VarDR1;t VarDR2;t . (7)
retailers mean demand. Note that Chen et al. (2000)
exhibit that the lower bound describes the behavior of the Note thatbecause of the zero lead time assump-
system accurately even in cases where z60. The relation- tionthe overall order-up-to level in (7) corresponds to
ship (6) shows that the increase in variability is a the inventory level ( stock on-handbackorders). How-
decreasing function of N, the number of observations N ever, when the wholesaler calculates the order-up-to level
PJ
used to estimate the mean and the variance of the based on the aggregated demand, j 1DRj,t, the whole-
retailers demand. salers order-up-to level (target stock level) in period

4. The bullwhip effect in supply networks

Chen et al. (2000) pointed out that the approach

presented above does not capture many of the complex- Retailer1
ities involved in real world supply chains. In the following,
we will extend this approach to account for the typical
Manufacturer
Manufacturer Wholesaler
Wholesaler
network structure of real supply chains. It will be shown
that the bullwhip effect may be overestimated if just a Retailer2
simple supply chain is assumed.
In practice, supply chains often exhibit a network
structure comprising of geographically dispersed produc- Fig. 3. The three-stage supply network.
 ARTICLE IN PRESS
316 E. Sucky / Int. J. Production Economics 118 (2009) 311322

t is given by demand can be calculated as

0 1 v
0 1
u q q2
X
2 u X 2
u VarDR1;t  VarDR2;t
~yt E@ A t
DRj;t z Var @ DRj;t A
j1 j1
0 1
X 2 q q2
0 1 @
pVar DRj;t Ap VarDR1;t VarDR2;t . (11)
X
2
E@ DRj;t A j1

j1
q The upper bound in (11) reveals a typical risk pooling
z VarDR1;t VarDR2;t 2CovDR1;t ; DR2;t , (8) effect in supply networks: in case of a coefcient of
correlation rDo1 a lower safety stock level has to be held
where Cov(DR1,t,DR2,t) is the covariance between the
to satisfy the desired safety level. The resulting reduction
demand of retailer (R1) and the demand of retailer (R2).
of inventory is well known as the square root law, which
If the retailers demands are stochastically independent,
was proven mathematically by Maister (1976). Mathema-
the covariance is zero. Then, the order-up-to level in (8)
tically stated, the square root law says that total inventory
for the aggregated demand is given by
in a system is proportional to the square root of the
q
y~ t EDR1;t EDR2;t z VarDR1;t VarDR2;t number of warehouses in which a product is stocked; i.e.
p p consolidating warehouses will have lower stock levels.
It is easy to see that the total VarDR1;t VarDR2;t
p Zinn et al. (1989), Zinn et al. (1990), and Ronen (1990)
in (7) is greater than VarDR1;t VarDR2;t : measure the effect of inventory centralization on stock
q q q level. Tyagi and Das (1998) analyze the structure of
VarDR1;t VarDR2;t 4 VarDR1;t VarDR2;t inventory systems based on the square root law. Lee et al.
qq (1993) show that the effect resulting from the square
3VarDR1;t 2 VarDR1;t VarDR2;t
root law is one reason for centralizing inventories in
VarDR2;t 4VarDR1;t VarDR2;t supply chains. Lee and Billington (1992, 1993, 1995) as
qq
well as Lee et al. (1993) demonstrate how HP utilized this
32 VarDR1;t VarDR2;t 40. (9)
risk pooling effect by implementing a so-called design for
Thus, the order-up-to levels y~ t based on the aggregated localization strategy. The risk pooling effect based on the
demand are less than the overall order-up-to levels yt square root law is a special case of the well-known
yR1;t yR2;t based on an isolated manner. portfolio effect.
In the following, the retailers demands DR1,t and DR2,t
are statistically correlated with a coefcient of correlation 4.2. The forecasting technique regarding risk pooling
1orDo1. With
The wholesaler still uses the simple N-period moving
CovD ; D
R1;t R2;t
rD p p , average MA(N):
VarDR1;t VarDR2;t 0 1
X2
the order-up-to level y~ t in (8) can be rewritten as E@ DRj;t A EDR1;t EDR2;t
0 1 j1
X
2 !
y~ t E@ DRj;t A 1 X
t1 X
t1

j1 DR1;i DR2;i
r
N
qq itN
!
itN
!
z VarDR1;t VarDR2;t 2rD VarDR1;t VarDR2;t . 1 X
t1
1 X
t1
DR1;i DR2;i , (12)
(10) N itN
N itN

Note that the retailers demands are perfectly positively 0 1

correlated if rD 1. The demands are perfectly negatively X
2
Var@ DRj;t A VarDR1;t VarDR2;t
correlated if rD 1. If rD 1 the variance of the j1
aggregated demand is given by qq
qq 2rD VarDR1;t VarDR2;t
VarDR1;t VarDR2;t 2 VarDR1;t VarDR2;t 0
q q2 1@ Xt1
D  EDR1;t 2
VarDR1;t VarDR2;t . N itN R1;i

For a correlation coefcient rD 1 the variance of the X

t1
DR2;i  EDR2;t 2
aggregated demand is given by itN
qq v
u t1
VarDR1;t VarDR2;t  2 VarDR1;t VarDR2;t uX
2rD t DR1;i  EDR1;t 2
q q2 itN
VarDR1;t  VarDR2;t . v1
u t1
uX
t DR2;i  EDR2;t 2 A. (13)
With the correlation coefcient rD 2 f1; 1g the upper
itN
and lower bounds of the variance of the aggregated
 ARTICLE IN PRESS
E. Sucky / Int. J. Production Economics 118 (2009) 311322 317

The order quantity, q~ t , the wholesaler places to the demands in supply networks is less than without
manufacturer in period tregarding the risk pooling consideration of risk pooling.
effect presented aboveis given by q~ t y~ t  y~ t1
P2
j1 DRj;t . To quantify the increase in variability from the 5. Experiments and simulation
wholesaler to the manufacturer, we determine the
variance of the orders placed by the wholesaler. In order 5.1. An illustrative example
to calculate a lower bound on the variance of the
wholesalers orders we, once again, assume z 0: In the following, we will illustrate the reduction of the
bullwhip effect if risk pooling effects are present. In this
0 1 example, the wholesaler preserves orders of two retailers
  X 2
2 2 (R1) and (R2). In the initial situation, the wholesaler
~
Varq 2 Var@
1 DRj A
N N j1 determines his orders separately based on the individual
  demand of each retailer, i.e. without consideration of a
2 2
1 2 VarDR1 VarDR2 potential correlation between the retailers demands. In
N N
pp this case, the overall order-up-to level in period t is given
2rD VarDR1 VarDR2 : (14) by yt yR1;t yR2;t (see Eq. (7)). With qt qR1;t qR2;t the
orders of the wholesaler are calculated. The wholesaler
If the wholesalers calculation is based on isolated uses a 2-period moving average MA(2) and a safety factor
demands of both retailers, with Eq. (7) the demands (z value) of z 2.33 (representing the desired service
p level). The calculation was performed on a Microsoft
variance is given by VarDR1 VarDR2 2 VarDR1
p Excels spreadsheet. Table 2 shows the relevant data for
VarDR2 . Therefore, if the wholesaler calculates the retailer (R1), i.e. the data to calculate yR1;t EDR1;t
order quantities in an isolated manner, i.e. qR1;t p
z VarDR1;t and qR1;t yR1;t  yR1;t1 DR1;t1 . Table 3
yR1;t  yR1;t1 DR1;t1 , qR2;t yR2;t  yR2;t1 DR2;t1 and shows the relevant data p for retailer (R2), which are
qt qR1;t qR2;t , the lower bound on the variance of the
used for yR2;t EDR2;t z VarDR2;t and qR2;t yR2;t 
wholesalers overall orders is given by yR2;t1 DR2;t1 . Table 4 shows yt yR1;t yR2;t and
qt qR1;t qR2;t .
  Fig. 4 shows that the wholesalers orders vary even
2 2
Varq 1 2 VarDR1 VarDR2 more than the retailers periodical demands. Calculating
N N
pp the sample variances Varq and Var(D) for the whole
2 VarDR1 VarDR2 .
planning periodsee Section 3.3the bullwhip
effect without consideration risk pooling is given
by Varq=VarD 1997:22=197:69 10:1 or Varq 
Consequently, if the wholesaler considers possible risk
VarD 1799:53.
pooling effects while calculating the order quantity, in
Determining the order-up-to levels and the orders
case of correlation rDo1 the orders variance Varq~ is less
for the retailers demands in an isolated manner
than the orders variance Varq, i.e. without considering
neglects the correlation of the demands. Analyzing
risk pooling. However, if we now determine the lower
the retailers periodical demands reveals that the
bound on the increase in variability from the wholesaler
demands of retailer (R1) and retailer (R2) are statisti-
to the manufacturer, i.e. the bullwhip effect, the relative
cally correlated with a correlation coefcient of
increase in variability generally does not change
rD 0.5. Table 5 shows the relevant data and the
and is still given by Eq. (6) as Varq=VarDX
results y~ t and q~ t . The target stock level y~ t is calcu-
1 2=N 2=N2 . Yet, the total increase in variability
lated by using Eq. (10) and the order quantity results
within the supply network regarding risk pooling is less P
from q~ t y~ t  y~ t1 2j1 DRj;t . Note that the corre-
than without taking risk pooling into account. The
lation coefcient rD is dened for the whole time
absolute increase of the variances of the orders placed
series. Therefore, computing the estimated variance
by the wholesaler relative to the variances of the demands
of the retailers demand Var(Dt) in period t, based
is given by
on (13), may differ from the results, which are obtained
using the technique in Section 3.2. However, Fig. 5
0 1 0 1
X
2   X 2 shows that the bullwhip effect is reduced signicantly if
2 2
~  Var@
Varq DRj A Var@ DRj A the wholesaler considers the statistical correlation of the
N N2
j1 j1 retailers demands. Regarding the statistically correlated
 
2 2 demands of the retailers, the bullwhip effect results t
2 VarDR1 VarDR2
N N ~
o Varq=VarD 400:24=197:69 2:02 or Varq ~ 
pp VarD 202:55.
2rD VarDR1 VarDR2 :
(15)
5.2. A simulation study

Therefore, in case of rDo1 the absolute increase of the Besides the illustrative example presented above, we
variances of the wholesalers orders to the variances of the have implemented a simulation study. Within this study,
 318
Table 2
The wholesalers orders depending on the demand of retailer (R1)

Period 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

DR1,t 94.46 103.7 81.69 135.37 99.63 98.05 86.82 100.40 104.32 89.33 84.70 105.68 120.92 91.21 109.74 90.75 104.08 119.80 107.21 123.18 72.90
E(DR1,t) 99.05 92.67 108.53 117.50 98.84 92.43 93.61 102.36 96.82 87.01 95.19 113.30 106.07 100.47 100.24 97.41 111.94 113.51 115.20
Var(DR1,t) 21.12 120.52 720.40 319.29 0.63 31.55 46.13 3.84 56.17 5.37 110.08 58.07 220.62 85.78 90.09 44.37 61.84 39.64 63.77
yR1;t 109.76 118.25 171.07 159.14 100.69 105.52 109.43 106.93 114.29 92.41 119.63 131.05 140.67 122.05 122.36 112.93 130.26 128.18 133.80
qR1;t 90.18 188.19 87.70 39.60 91.65 104.31 101.81 96.69 62.82 132.90 132.34 100.83 91.12 91.06 94.65 137.13 105.13 128.81

E. Sucky / Int. J. Production Economics 118 (2009) 311322

ARTICLE IN PRESS
Table 3
The wholesalers orders depending on the demand of retailer (R2)

Period 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

DR2,t 99.08 89.55 123.32 72.48 72.69 92.61 107.71 76.74 90.49 104.21 87.74 99.04 104.23 83.22 89.31 116.53 102.40 90.59 90.79 78.75 128.18
E(DR2,t) 94.31 106.43 97.90 72.58 82.65 100.16 92.23 83.62 97.35 95.98 93.39 101.64 93.72 86.26 102.92 109.46 96.49 90.69 84.77
Var(DR2,t) 22.74 285.18 646.25 0.01 99.17 56.98 239.65 47.22 47.10 67.82 31.93 6.73 110.43 9.27 185.24 49.92 34.85 0.01 36.25
yR2;t 105.43 145.78 157.13 72.83 105.85 117.75 128.29 99.63 113.34 115.17 106.56 107.68 118.21 93.36 134.63 125.92 110.25 90.92 98.79
qR2;t 163.67 83.83 11.60 125.63 119.60 87.29 61.82 117.93 89.57 90.43 105.36 93.74 64.45 157.80 93.69 74.91 71.46 86.62
 ARTICLE IN PRESS
E. Sucky / Int. J. Production Economics 118 (2009) 311322 319

215.43
232.59
201.08
300

20
250

201.93

176.59
219.1
200
19

Quantity
198.00

212.04
240.51
150
18

Demand of the retailers

100 Wholesaler's orders
210.39
238.85
188.34

50
17

3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 20
256.99
248.86
206.47

Time
16

Fig. 4. Demand and order quantities without consideration of risk

pooling.
207.28

155.57
215.41
15

258.88
199.04

194.57
14

we have simulated the demand for time series of 1000

periods, with E(DR1) E(DR2) 100 and Var(DR1)
174.43
238.73

Var(DR2) 400 over all periods and a safety factor

237.7

(z value) of z 1.645 (representing the desired service

13

level of 95%). The correlation coefcient varies between

223.33

rDA[0.9;0.9] in steps of 0.1; i.e. we have analyzed 19

225.15
226.19

scenarios for the correlation coefcient. Each scenario was

12

calculated 100 times.

Fig. 6 shows that the average bullwhip effect for each
152.39
207.58
204.72

scenario varies between 2.5 and 3; i.e. 2:5o

11

~
Varq=VarDo3:0. As Chen et al. (2000) noted, the lower
bound (6) accurately describes the behavior of the system
172.44
227.63
214.62

even in cases where z60.

10

In our simulation study we use a 2-period moving

average MA(2), i.e. N 2. Therefore, the resulting
193.54

163.63
206.56

~
lower bound is given by Varq=VarDX2:5. The relative
9

increase in variability does not vary depending on

the correlation coefcient. It is easy to see that the
194.81
237.72
191.6

bullwhip effect exists denitely. However, Fig. 7 shows

the variance of the demand Var(D) and the variance of
8

the orders placed by the wholesaler Var(q) depending

211.25
223.27
177.14

on the value of the correlation coefcient. It is shown

that both the variance of the demand Var(D) and the
7

variance of the orders placed by the wholesaler

The wholesalers orders without consideration risk pooling

194.52

165.23
206.85

Var(q) decrease for lower values of the correlation

coefcient. Measuring the bullwhip effect by the
6

absolute difference Var(q)Var(D), the reduction of the

190.66
231.97

bullwhip effect for correlated demands is considerable

76.1

(see Fig. 8).

5

172.33

272.02
328.2
4

6. The bullwhip effect in supply networks with multiple

retailers
207.85

253.85
264.03
3

In the preceding section, we have simplied our

analysis to the case of two retailers at one stage of
yt yR1;t yR2;t
qt qR1;t qR2;t

the supply network. In reality, a wholesaler often

DR1,t+DR2,t

supplies more than two retailers. If the wholesaler

Table 4

Period

lls the demands of J retailers, with J42, the variance

of the aggregated demand can be calculated as (see Mood
 ARTICLE IN PRESS
320 E. Sucky / Int. J. Production Economics 118 (2009) 311322
3.86
199.96

199.93
201.08

216.76

300
20

250
201.93

218.75
191.80
38.41
204.19

200
19

Quantity
198.00
208.43
3.84
224.95
212.45

150
18

100
Demand of the retailers
222.89
210.39
206.88

198.03
0.16

Wholesaler's orders
50
17

0
16.96

232.82
231.33
206.47
203.16

3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
16

Time
207.28

151.45
205.79
174.22
186.74

Fig. 5. Demand and order quantities with consideration of risk pooling.

15

199.79
199.04

643.22
230.61
174.47

4.0
14

3.5
245.95
174.43
214.94
104.35
230.57

Var (q)/Var (D)

13

3.0
188.58

209.77
260.57

214.37
225.15

2.5
12

2.0
182.99
111.36

161.62
204.72

200.13

1.5
11

1.0
172.44

0.39
210.95
204.23
194.18

-0.9 -0.7 -0.5 -0.3 -0.1 0.1 0.3 0.5 0.7 0.9
10

Correlation coefficient D
200.26
193.54
185.98
77.99

177.92

Fig. 6. Bullwhip effect depending on the value of the correlation

9

coefcient.
194.81
185.83

185.88
217.15
75.5
8

4500
192.59

208.42
3.73
177.14

199.11

4000
7

3500
3000
194.52
181.49

203.83
162.89
84.02

Var (q)

Variance
2500
6

2000
Var (D)
136.54
190.66
190.09

231.60
315.47

1500
5

1000
The wholesalers orders with consideration risk pooling

500
267.38
241.54
172.33
206.43
2.01

0
4

-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1

34.92
207.85

233.69
234.41 Correlation coefficient D
199.10
3

Fig. 7. Variances of the demands and the orders depending on the

204.28
0.03
193.37
205.01
correlation coefcient.
2

193.2
1
et al., 1974)

193.5
0 1
J
X J
X J
X X
K

0
Var@ DRj;t A VarDRj;t 2 CovDRj;t ; DRk;t .

Table 5

Var(Dt)
Period
j1 j1 j1 kj1

E(Dt)
(16)

Dt

q~ t
y~ t
 ARTICLE IN PRESS
E. Sucky / Int. J. Production Economics 118 (2009) 311322 321

4000 coefcients less than one could strongly reduce the

3500 bullwhip effect.

3000
Var (q) - Var (D)

2500 7. Conclusion
2000
1500 By assuming a three-stage supply chain consisting of a
single retailer, a single wholesaler, and a single manufac-
1000
turer relevant risk pooling effects associated with the
500 network structure of supply chains are neglected. Based
0 on the approaches by Chen et al. (1999, 2000) we have
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 shown that the bullwhip effect may be overestimated if
Correlation coefficient D just a supply chain is assumed and risk pooling effects in
supply networks can be utilized. If we take into the
Fig. 8. The absolute difference Var(q)Var(D) depending on the consideration that in practice forecasting methods super-
correlation coefcient. ior to the simple N-period moving average are used and
regarding the ndings of Cachon et al. (2005), we can
The Eq. (16) can be expressed using the covariance conclude that the bullwhip effect is present; however, the
matrix: bullwhip effect is not commonplace. Finally, referring to

2 3
EDR1;t  EDR1;t DR1;t  EDR1;t   EDR1;t  EDR1;t DRJ;t  EDRJ;t 
6 ED EDR2;t  EDR2;t DRJ;t  EDRJ;t  7
6 R2;t  EDR1;t DR1;t  EDR1;t   7
6 7
6 .. .. .. 7
6 . . . 7
4 5
EDRJ;t  EDRJ;t DR1;t  EDR1;t   EDRJ;t  EDRJ;t DRJ;t  EDRJ;t 
2 3
VarDR1;t CovDR1;t ; DR2;t  CovDR1;t ; DRJ;t
6 CovD ; D VarDR2;t  CovDR2;t ; DRJ;t 7
6 R2;t R1;t 7
6 7
6 .. ... .. ... 7. (17)
6 . . 7
4 5
CovDRJ;t ; DR1;t CovDRJ;t ; DR2;t  VarDRJ;t

In case of J retailers (J42), the wholesaler may still use Dejonckheere et al. (2003), we come to the logical
the N-period moving average MA(N) determining conclusion that order-up-to systems usually result in the
P P
E Jj1 DRj;t and Var Jj1 DRj;t : bullwhip effect, but the strength of the effect depends on
0 1 ! the statistical correlation of the regarded demands.
X J X
t1 X
t1 X
t1
1
E@ DRj;t A DR1;i DR2;i    DRJ;i ,
j1
N itN itN itN
Acknowledgement
(18)

0 1 The author would like to thank the anonymous

X J   X
t1 referees who have provided invaluable critique.
1
Var@ DRj;t A DR1;i  EDR1;t 2
j1
N itN
References
X
t1
2
DR2;i  EDR2;t   
itN Anderson, E., Fine, C., Parker, G., 2000. Upstream volatility in the supply
chain: The machine tool industry as a case study. Production and
X
t1
Operations Management 9, 239261.
DRJ;i  EDRJ;t 2 Cachon, G.P., Randall, T., Schmidt, G.M., 2005. In search of the bullwhip
itN effect. Manufacturing & Service Operations Management 9 (4),
0 v
J u t1 457479.
X X
K uX
2 @r t DRj;i  EDRj;t 2 Carlsson, C., Fuller, R., 2000. A fuzzy approach to the bullwhip effect. In:
j;k Cybernetics and Systems 2000, Proceedings of the 15th European
j1 kj1 itN
v11 Meeting on Cybernetics and System Research, Vienna.
u t1 Carlsson, C., Fuller, R., 2001. Reducing the bullwhip effect by means of
uX
t DRk;i  EDRk;t 2 AA: (19) intelligent, soft computing methods. In: Proceedings of the 34th
Hawaii International Conference on System Sciences, pp. 110.
itN
Chen, F., Drezner, Z., Ryan, J.K., Simchi-Levi, D., 1999. The bullwhip effect:
managerial insights on the impact of forecasting and information on
With (18) and (19) the order quantity qt, the wholesaler variability in a supply chain. In: Tayur, S., Ganeshan, R., Magazine, M.
places to the manufacturer in each period t, can be (Eds.), Quantitative Models for Supply Chain Management, Boston,
calculated. Finally, using the difference Var(q)Var(D), the Dordrecht, London, pp. 419439.
Chen, F., Drezner, Z., Ryan, J.K., Simchi-Levi, D., 2000. Quantifying the
bullwhip effect can be determined in case of J42 retailers. bullwhip effect in a simple supply chain: The impact of forecasting,
Then, risk pooling effects resulting from correlation lead times, and information. Management Science 46 (3), 436443.
 ARTICLE IN PRESS
322 E. Sucky / Int. J. Production Economics 118 (2009) 311322

Dejonckheere, J., Disney, S.M., Lambrecht, M.R., Towill, D.R., 2003. Mason-Jones, R., Towill, D.R., 2000. Coping with uncertainty: Reducing
Measuring and avoiding the bullwhip effect: A control theoretic Bullwhip Behaviour in global supply chains. Supply Chain Forum
approach. European Journal of Operational Research 147, 567590. (1), 4045.
Disney, S.M., Towill, D.R., 2003. Vendor-managed inventory and bullwhip Metters, R., 1997. Quantifying the bullwhip effect in supply chains.
reduction in a two-level supply chain. International Journal of Journal of Operations Management 15, 89100.
Operations & Production Management 23 (6), 625651. Mood, A.M., Graybill, F.A., Boes, D.C., 1974. Introduction to the Theory of
Forrester, J.W., 1958. Industrial dynamics, A major breakthrough for Statistics. New York.
decision makers. Harvard Business Review JulyAugust, 6796. Moyaux, T., Chaib-draa, B., DAmours, S., 2003. Multi-agent coordinatiion
Forrester, J.W., 1961. Industrial Dynamics, New York, London. based on tokens: reduction of the bullwhip effect in a forest supply
Geary, S., Disney, S.M., Towill, D.R., 2006. On bullwhip in supply chain. In: Proceedings of the AAMAS03, Melbourne, Australia.
chainsHistorical review, present practice and expected future Moyaux, T., Chaib-draa, B., DAmours, S., 2007. Information sharing as a
impact. International Journal of Production Economics 101 (1), 218. coordination mechanism for reducing the bullwhip effect in a supply
Hammond, J., 1994. Barilla SpA (a), Harvard Business School case #9- chain. IEEE Transactions on Systems, Man, and Cybernetics 37 (3),
694-046. 396409.
Ingalls, R.C., Foote, B.L., Krishnamoorthy, A., 2005. Reducing the bullwhip Nienhaus, J., Ziegenbein, A., Schonsleben, P., 2006. How human
effect in supply chains with control-based forecasting. International behaviour amplies the bullwhip effectA study based on the beer
Journal of Simulation & Process Modelling 1 (1/2), 90110. distribution game online. Production, Planning & Control 17 (6),
Johnson, L.A., Montgomery, D.C., 1974. Operations Research in Production 547557.
Planning, Scheduling, and Inventory Control. New York, London. Ronen, D., 1990. Inventory centralization/decentralizationThe
Kahn, J., 1987. Inventories and the volatility of production. The American square root law revisited again. Journal of Business Logistics 11 (2),
Economic Review 77 (4), 667677. 129138.
Lee, H.L., Billington, C., 1992. Managing supply chain inventory: Pitfalls Silver, E.A., Pyke, D.F., Peterson, R., 1998. Inventory Management and
and opportunities. Sloan Management Review 33 (3), 6573. Production Planning and Scheduling. Third Edition, New York,
Lee, H.L., Billington, C., 1993. Material management in decentralized Singapore, Toronto.
supply chains. Operations Research 41 (5), 835847. Simchi-Levi, D., Kaminsky, P., Simchi-Levi, E., 2000. Designing and
Lee, H.L., Billington, C., 1995. The evolution of supply-chain-management Managing the Supply ChainConcepts, Strategies, and Case Studies.
models and practice at HewlettPackard. Interfaces 23 (4), 4263. Boston, New York, Sydney.
Lee, H.L., Billington, C., Carter, B., 1993. HewlettPackard gains control of Sterman, J., 1989. Modeling managerial behavior: Misperceptions of
inventory and service through design for localization. Interfaces 23 feedback in a dynamic decision making experiment. Management
(4), 111. Science 35, 321339.
Lee, H.L., Padmanabhan, V., Whang, S., 1997a. The bullwhip effect in Terwiesch, C., Ren, T., Ho, H., Cohen, M., 2005. Forecast sharing in the
supply chains. Sloan Management Review, 93102. semiconductor equipment supply chain. Management Science 51,
Lee, H.L., Padmanabhan, V., Whang, S., 1997b. Information distortion in a 208220.
supply chain: The bullwhip effect. Management Science 43 (4), 546558. Tyagi, R., Das, Ch., 1998. Extension of the square-root law for safety stock
Lee, H.L., Padmanabhan, V., Whang, S., 2004. Comments on information to demands with unequal variances. Journal of Business Logistics 19
distortion in a supply chain: The bullwhip effectThe bullwhip (2), 197203.
effect: Reections. Management Science 50 (12), 18871893. Zhou, L., Disney, S.M., 2006. Bullwhip and inventory variance in a closed
Maister, D.H., 1976. Centralization of inventories and the square root law. loop supply chain. OR Spectrum 28, 127149.
International Journal of Physical Distribution 6 (3), 124134. Zinn, W., Levy, M., Bowersox, D.J., 1989. Measuring the effect of inventory
Makajic-Nikolic, D., Panic, B., Vujosevic, M., 2004. Bullwhip effect and centralization/decentralization on aggregate safety stock: The square
supply chain modelling and analysis using CPN Tools. In: Jensen, K. root law revisited. Journal of Business Logistics 10 (1), 114.
(Ed.), Proceedings of the Fifth Workshop and Tutorial on Practical Zinn, W., Levy, M., Bowersox, D.J., 1990. On assumed assumptions and the
Use of Coloured Petri Nets and the CPN Tools, Aarhus, Denmark, inventory centralization/decentralization issue. Journal of Business
pp. 219234. Logistics 11 (2), 139142.