Sei sulla pagina 1di 13

Production, Manufacturing and Logistics

Analysis of compound bullwhip effect causes


Xiaolong Zhang

, Gerard J. Burke
1
Department of Finance and Quantitative Analysis, College of Business Administration , Georgia Southern University, Statesboro, GA 30460, United States
a r t i c l e i n f o
Article history:
Received 9 April 2009
Accepted 22 September 2010
Available online 29 September 2010
Keywords:
Bullwhip effect
Forecasting
Supply chain
Inventory control
a b s t r a c t
This research investigates compound causes of the bullwhip effect (BWE) by considering an inventory
system with multiple price-sensitive demand streams. Joint price and demand dynamics are captured
by a vector time-series process that incorporates the stochastic co-movements in price and demand.
We study two BWE measures, one for each demand stream individually and one for the aggregated
demand. We show that demand parameters including demand autocorrelation, cross-correlation, and
price sensitivity serve as root causes of the BWE. We prove that the impact of these parameters on the
BWE can be additively decomposed. Conditions are established under which a pair of simultaneous com-
pound causes may attenuate or dampen the BWE. When demand streams are aggregated, we derive a
pooling factor that quanties the impact of demand aggregation on order stability. When positive, the
pooling factor corresponds to a synergy effect that captures the gain in the stability of the pooled orders.
Conditions for the existence of the synergy effect are obtained for several special cases involving a zero
leadtime. We also discuss how our analytical ndings can be managerially applied to bullwhip mitigation
strategies.
2010 Elsevier B.V. All rights reserved.
1. Introduction
Industrial dynamics seeks to quantitatively predict systemic or industrial effects caused by dynamic rm-level decision making (For-
rester, 1958, 1961). One such systemic effect of contemporary interest is the bullwhip effect (BWE). When it exists, the BWE propagates
and amplies demand and order variability upstream, which can create operational instability throughout supply chains. A seminal paper
by Lee et al. (1997a) investigates several causes for the BWE including serial forecasting techniques, inventory control policies, and price
promotions. It is often, however, not a single cause that contributes to the magnitude of the BWE. We propose in this paper an analytical
framework that incorporates multiple compounding causes and explores their interactions in determining the BWE.
Our research is partly motivated by the experiences of Procter & Gamble and Cisco Systems, both suffered from the BWE due to com-
pound causes. P&Gs experience involves disposable diapers, a product with relatively stationary consumer demand. Serial forecasting
techniques, inventory control policies, and pricing promotions (Lee et al., 1997a,b) have in concert attributed to boom and bust replenish-
ment cycles more volatile than those propagated by each individual cause alone. In Ciscos case, the BWE contributed to a $2.25 billion
dollar write-off for its network hardware components in 2001, a nancial asco most notably caused by inated serial demand forecasts
and multiple orders for the same demand due to network gear component scarcity (i.e. ration gaming) (Barrett, 2001).
These rm specic instances illustrate the multi-interactive-cause nature of the BWE. Previous research, however, has focused on quan-
tifying BWE one cause at a time. Our research considers compound BWE causes stemming from demand dynamics and price sensitivity and
seeks to quantify their separate and aggregate inuences on order variability.
Theoretical research on quantifying the BWE size due to an individual cause has relied largely on either time-series or control-theory
methodologies. Lee et al. (1997a) originated the time-domain approach in studying forecasting as a cause for the BWE when demand is a
rst-order autoregressive (AR) process. Subsequent papers expanded their work by allowing for different forecasting methods (Chen
et al., 2000; Zhang, 2004a), more general time-series demand processes (Graves, 1999; Aviv, 2003; Zhang, 2004b; Gilbert, 2005), or different
information sharing settings (Lee et al., 2000; Raghunathan, 2001; Gaur et al., 2005). In addition, frequency-domain transfer functions have
been applied to link the input demand with output order process for a given inventory control policy when quantifying the BWE (Ouyang
and Daganzo, 2006; Ouyang, 2007; Ouyang and Li, 2010; Jaksic and Rusjan, 2008). While a time-domain approach facilitates the exploration
0377-2217/$ - see front matter 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2010.09.030

Corresponding author. Tel.: +1 912 478 5741; fax: +1 912 478 1835.
E-mail addresses: xzhang@georgiasouthern.edu (X. Zhang), gburke@georgiasouthern.edu (G.J. Burke).
1
Tel.: +1 912 478 1495; fax: +1 912 478 1835.
European Journal of Operational Research 210 (2011) 514526
Contents lists available at ScienceDirect
European Journal of Operational Research
j our nal homepage: www. el sevi er . com/ l ocat e/ ej or
of time-series structure of the order process for a given family of demand processes and inventory control policy, the frequency-domain
approach provides a convenient means for studying the cyclical behavior in the demand and order processes. Control theoretic approaches
are implemented by Dejonckheere et al. (2003, 2004) to measure the BWE (2003) and to evaluate the effectiveness of demand information
transparency (2004). Papers by Lee et al. (2004), Kouvelis et al. (2006), Disney and Lambrecht (2007) provide more in-depth reviews of the-
oretical literature on the BWE.
Besides theoretical efforts, attempts are also made to validate the BWE in experimental settings or with empirical data. Variations of the
popular beer-game of Sterman (1989) have repeatedly shown that the BWE shows up even when demand uncertainty is completely
eliminated (Wu and Katok, 2006). Experimental results often contradict the expected smoothing effects from inventory and aggregation
of demand information. Likewise, a vast number of research papers in empirical economics contradict production smoothing. Cachon et
al. (2007) provides an excellent review of empirical economic research in many industries that nds production to be more variable than
sales. Cost shocks in production, batch production and distribution, and auto-regressive demand are cited explanations for the lack of sup-
port for production smoothing. They note that various factors inuencing production can occur simultaneously: e.g., it is possible that a
rm faces increasing marginal costs and xed ordering costs and positively correlated demand. Hence, whether or not a rm smoothes
production relative to sales depends on the relative importance of these factors. Cachon et al. (2007) empirically searches for the bullwhip
in industry level retail, wholesale, and manufacturing Census data. They fail to nd strong support for the BWE and predictable seasonality
in demand was found to mitigate the BWE signicantly.
In the spirit of Cachon et al. (2007), our research is aimed at investigating, at a rm-level, the simultaneous impact of multiple causes on
the BWE and the manner in which their impact is compounded. We continue theoretical investigations of the BWE using a time-domain
approach; our results, however, are both prescriptive and descriptive. We differ from previous research in three aspects. First, we incor-
porate price dynamics into multiple demand streams. Price dynamics are captured by an AR(1) process and multiple demand streams
are described by a joint time-series process with its own separate dynamics. This allows us, for the rst time, to examine and quantify
the BWE when price uctuates and multiple demands interact and correlate. Second, we focus on how multiple causes in our theoretical
model are compounded in determining the overall BWE. Our approach facilitates decomposing the overall BWE into components that have
assignable causes. Lastly, we propose an aggregate BWE measure to explore the risk-pooling opportunities and synergy across demand
streams that can be used to combat the BWE.
The BWE elevates the risk that companies face in inventory planning. Our research ndings can be viewed as providing operational or
natural hedges against demand risks when multiple BWE causes interact in a replenishment system. From this perspective our work pro-
vides a bridge between BWE time-series research and operational hedging research areas in supply chain management. Boyabatli and Tok-
tay (2004) provide a thorough review of operational hedging literature. Other related papers are Mieghem (2007), Burke et al. (2007),
Corbett and Rajaram (2006), Harrison and Van Mieghem (1999). In our paper, operational hedging opportunities arise from downstream
demand diversity (a real option) with price shocks. In this sense we illustrate that market bundling can inherently mitigate BWE-driven
risk. We also explicitly consider the compounding or combined effects of multiple BWE causes on a replenishment system. Furthermore,
we quantify the degree to which these simultaneous inuences dampen or amplify actual market demand variability and show that multi-
ple BWE sources can actually reduce demand variability upstream in a supply chain.
In the next section, we present a multiple demand stream replenishment model and dene two types of BWE. In Section 3, we describe a
joint demand model incorporating price dynamics. In Section 4, we use demand parameters to serve as compounding factors and explore
how they impact the BWE separately and in aggregate. In Section 5, we describe and report ndings from numerical experimentation that
investigates likely parametric ranges of compounding effects. Finally, in Section 6, we conclude the paper, provide directions for future re-
search, and discuss managerial implications.
2. Replenishment model
Our replenishment model represents an inventory system that supplies two demand streams. It is triadic: we have one supplier lling
orders to meet two demand streams that may represent two separate retailer demands, product lines, customer groups, or geographical
market segments. For convenience, we will refer to the three parties as two retailers and a wholesaler. We assume that the only demands
the wholesaler faces are orders from these two retailers. The wholesaler maintains its own inventory and lls both retailers orders with a
xed lead time L. The wholesaler orders from an outside source that has ample capacity and replenishes the wholesaler with a lead time of
L
0
. The supply system is decoupled with decentralized control. As such, the inventory control for each demand stream can be treated sep-
arately. Our supply chain setting has been commonly used in the time-domain BWE literature (Lee et al., 2000; Raghunathan, 2001; Aviv,
2001, 2003; Lee et al., 1997a; Chen et al., 2000).
Inventory related costs are assumed proportional and there are no xed ordering costs in the supply chain. Furthermore, all retail short-
ages are back-ordered. We follow this sequence of events in an ordering cycle t: (1) at the beginning of t, inventories are reviewed, retailers
orders made during t L are received, and the wholesalers orders made during t L
0
are received; (2) retailers place their orders for the
period and are transmitted and processed instantly; (3) the wholesaler places its order; and (4) demand is realized.
Given our decoupled supply system, a reasonable control policy is a factored base-stock type with time-varying order-up-to levels in
which each retailer controls its own inventory in a decentralized manner. For retailer i, let d
i, t
be the demand during period t, y
i, t
be the
order-up-to level, and q
i, t
be the corresponding orders made in period t. Conservation of ow implies that
q
i;t
= y
i;t
y
i;t1
d
i;t1
: (1)
Further, we dene lead-time demand as D
i
[t; t L[ =

L
s=0
d
i;ts
and let

D
i
[t; t L[ be the MMSE (minimum mean-square error) forecast based
on observed demand histories up to and including t. As shown in Zipkin (2000, pp. 415420), for normally distributed demand, an approx-
imately optimal order-up-to level can be calculated as the sum of forecasted lead-time demand and safety stock:
y
i;t
=

D
i
[t; t L[ f^ r
i;t
; (2)
X. Zhang, G.J. Burke / European Journal of Operational Research 210 (2011) 514526 515
where f is a standard normal score derived from desired service level, and ^ r
i;t
=

V(D
t
[t; t L[

D
t
[t; t L[)
_
is the standard deviation of lead-
time demand forecast error. The order-up-to levels consist of an anticipation stock kept to meet the expected demand during lead time and a
safety stock for hedging against unexpected demand. The factored replenishment policy for multiple demand streams specied in Eq. (2) has
been studied by Veinott (1965). A later paper by Johnson and Thompson (1975) showed specically that under general conditions a similar
policy is optimal for general autoregressive and moving average demand processes. These results provide rationale behind our control policy,
which can be considered an approximation to the optimal policy.
Combining Eqs. (1) and (2), we have:
q
i;t
=

D
i
[t; t L[

D
i
[t 1; t L 1[ f(^ r
i;t
^ r
i;t1
) d
i;t1
: (3)
Hence, orders are made in each period to replace demand and compensate for policy driven changes in levels of anticipation and safety
stocks.
2
In all cases we consider, the standard deviation of lead-time forecast error ^ r
i;t
remains constant over time and the order quantity
in Eq. (3) simplies to:
q
i;t
=

D
i
[t; t L[

D
i
[t 1; t L 1[ d
i;t1
: (4)
We consider two types of BWE measures: one is based on the demand and order stream of an individual retailer and the other is con-
structed to measure the BWE of pooled retail demands and orders. For simplicity, both measures are based on differences between order
variance and demand variance as dened below. By using difference measures rather than the conventional ratio measures, our results can
be more compactly expressed and better interpreted.
Denition 1. A separate BWE exists if V(q
i,t
) > V(d
i,t
) and the size of the effect is measured by their difference M
i
= V(q
i,t
) V(d
i,t
).
Denition 2. An aggregate BWE exists if V(q
1,t
+ q
2,t
) > V(d
1,t
+ d
2,t
) and M = V(q
1;t
q
2;t
) V(d
1;t
d
2;t
) measures the size of the BWE.
In both denitions, when a BWE does not exist, we refer to it as a zero BWE if the variance of the order is equal to the variance of the
demand, or as a smoothing effect if the variance of the order is less than the variance of the demand. Expanding the variance of the sum in
Denition 2, the aggregate BWE measure can be expressed as
M = M
1
M
2
2[COV(q
1;t
; q
2;t
) COV(d
1;t
; d
2;t
)[; (5)
the sum of the two separate BWE measures plus an adjustment to account for the covariance between the order streams and the demand
streams. In our subsequent discussions, we refer to this adjustment as the pooling factor. When the pooling factor is negative, M is less than
the sum of the separate M
i
. There is smoothing effect from demand diversication. Hence, pooling orders dampens the aggregate BWE, indi-
cating a synergy effect between the two demand streams as well as between the two order streams. On the other hand, when the pooling
factor is positive, the aggregated BWE is larger than the sum of the separate BWE, indicating an effect opposite to the synergy effect.
3. The demand model
We explicitly incorporate price dynamics in our demand specication. The role of price change in the stability of supply chains has been
studied as reverse BWE (Ozelkan and Cakanyldrm, 2009). Their focus is on the price variance magnication when an upstream installa-
tion such as a wholesaler enters a price agreement with a downstream installation such as a retailer. The price agreement between the two
parties makes retail price endogenous and controllable. Similarly, Federgruen and Heching (1999) models a monopoly market in which a
rm exerts complete control over the market price. Instead, we consider a market setting in which both retailers sell on a perfectly com-
petitive market and exert no control over the market clearing price. Thus, the market price evolution is exogenously determined by the
forces of overall market demand and supply. This price determination scenario is typical for products traded through competitive supply
markets where global shocks may trigger industry-wide, economy-wide or world-wide price dynamics. We describe price dynamics with a
rst-order autoregressive process
p
t
= a
p
kp
t1
e
p;t
; (6)
where the error term e
p,t
is i.i.d. with a standard deviation of r
p
. Our price dynamic can be considered a discrete version of the short-term
component in the two-factor model of commodity prices proposed by Schwartz and Smith (2000) who synthesize the random walk and
mean-reverting commodity pricing models. Our AR(1) pricing assumption is consistent with their approach. In a mean-centered form, we
can equivalently express the price dynamics as
z
p;t
= kz
p;t1
e
p;t
;
z
p;t
= p
t
l
p
; l
p
= a
p
1 k ( )
1
:
(7)
Given the market price dynamics, we assume that the demand streams facing the two retailers are jointly determined by a linear com-
bination of demand inertia, interaction, and price changes
d
1;t
= a
1
q
11
d
1;t1
q
12
d
2;t1
b
1p
p
t
e
1;t
;
d
2;t
= a
2
q
21
d
1;t1
q
22
d
2;t1
b
2p
p
t
e
2;t
;
(8)
where a
1
and a
2
are constant intercept terms that determine the means of d
1,t
and d
2,t
; coefcients q
11
and q
22
determine autocorrelation due
to demand inertia; coefcients q
12
and q
21
describe the interaction between two demand streams; coefcients b
ip
describe the demand
2
The prescribed ordering policy allows the order quantity to be negative. If the probability of negative order quantity is negligible, however, our results remain valid as a close
approximation. Earlier studies (Lee et al., 1997a,b; Chen et al., 2000; Graves, 1999) maintained this feature to simplify their analysis.
516 X. Zhang, G.J. Burke / European Journal of Operational Research 210 (2011) 514526
responsiveness to market price changes over time. Random error terms, e
1,t
, e
2,t
, and e
p,t
, are jointly distributed normal. In vector form,
v
t
= [e
1,t
, e
2,t
, e
p,t
]
T
have the following covariance structure,
E v
t
v
T
t
/
_ _
= 0; \t t
/
;
E v
t
v
T
t
_ _
= V =
r
11
r
12
r
1p
r
12
r
22
r
2p
r
1p
r
2p
r
pp
_

_
_

_
:
(9)
Errors are assumed independent across time but correlated contemporaneously in our joint demand model. If we interpret e
p,t
as the effect on
price from overall market shocks, the above correlation structure implies that the overall market shocks captured by e
p,t
are correlated with
the demand shocks specic to the retailers. For analytical convenience, we express the demand process in a mean-centered form by subtract-
ing the unconditional (stationary) mean from its respective demand. Denote l
1
and l
2
as the unconditional mean demand for retailers 1 and
2 and let z
i, t
= d
i, t
l
i
for i = 1, 2. The joint demand process incorporating market price effects can be rewritten in mean-centered form
z
1;t
= q
11
z
1;t1
q
12
z
2;t1
b
1p
z
p;t
e
1;t
;
z
2;t
= q
21
z
1;t1
q
22
z
2;t1
b
2p
z
p;t
e
2;t
;
(10)
where
l
1
=
(1q
22
)(a
1
b
1p
l
p
)q
12
(a
2
b
2p
l
p
)
(1q
11
)(1q
22
)q
12
q
21
; l
2
=
(1q
11
)(a
2
b
2p
l
p
)q
21
(a
1
b
1p
l
p
)
(1q
11
)(1q
22
)q
12
q
21
:
4. Quantifying the BWE
We quantify the separate and aggregate BWE when both retailers place orders according to Eq. (4) to meet their demand in Eq. (8). For
demand forecasting, we assume that the joint demand process is known to all parties in the supply system and they have access to the
demand history of both demand streams. In order to isolate the incremental impact of a compound cause and to develop the analytical
tools for the general case, we begin with a benchmark case in which we eliminate price changes as a compound cause. We then examine
the BWE under the full demand specication with price changes in effect.
4.1. The benchmark: reduced demand process
To establish the benchmark case, we restrict price-related demand parameters to be zero, i.e.,
b
1p
= b
2p
= r
1p
= r
2p
= 0:
Doing so excludes inuences of price changes and still allows us to investigate impacts on the BWE from following two compounding causes
beyond demand autocorrelation: (1) deterministic interaction between demand streams captured by q
12
and q
21
and (2) demand cross cor-
relation r
12
. The joint demand process now simplies into a VAR(1) (rst-order vector autoregressive) process:
d
1;t
= a
1
q
11
d
1;t1
q
12
d
2;t1
e
1;t
; d
2;t
= a
2
q
21
d
1;t1
q
22
d
2;t1
e
2;t
:
In matrix notation, let d
t
= [d
1,t
, d
2,t
]
T
, l = [l
1
,l
2
]
T
, z
t
= d
t
l = [z
1,t
, z
2,t
]
T
, and e
t
= [e
1,t
, e
2,t
]
T
, then we can express the restricted joint demand
as
d
t
= a qd
t1
e
t
; z
t
= qz
t1
e
t
;
q =
q
11
q
12
q
21
q
22
_ _
; V(e
t
) = R =
r
11
r
12
r
12
r
22
_ _
; COV e
t
; e
t
/ ( ) = 0 \t t
/
:
(11)
A vector ARMA time series model has been frequently used to study the joint dynamics between sales and advertising (Hanssens et al., 2001;
Dekimpe and Hanssens, 1995; Dekimpe et al., 1998), among nancial time-series (Tsay, 2002), and among macroeconomic measurements
(Ltkepohl, 2005. Zhang and Zhao, 2010) employed the same demand process in (11) to study the value of acquiring demand information
from another demand stream for two parallel serial supply chains. They provide economic and marketing motivations that may underlie
such a demand process. This restricted demand process provides a non-trivial starting point for us to investigate important causes for
BWE that have yet been thoroughly explored in the literature.
To arrive at our explicit analytical results that quantify the BWE, we take four steps. We rst present a well-known formula for calcu-
lating the covariance matrix of restricted joint demand and then calculate each retailers MMSE forecast. From these expressions, we can
derive respective order quantities. As a last step, the calculation of the variance of the order quantities leads us to our separate and aggre-
gate BWE measures.
The covariance matrix of joint demand V(d
t
) can be easily obtained from YuleWalker equations for a VAR(1) time series process. Ltke-
pohl (2005) contains the technical details on deriving the covariance matrix for general vector ARMA processes, and it is given by
V(d
t
) = qV(d
t
)q
T
R: (12)
Let the symbol represent the Kronecker product of two matrices and vec() be the vector operator that stacks the columns of a matrix to
form a single vector. If we apply the vec( ) operator to (12), we have
vec[V(d
t
)[ = (I q q)
1
vec(R): (13)
X. Zhang, G.J. Burke / European Journal of Operational Research 210 (2011) 514526 517
Next we characterize each retailers MMSE forecast and respective order quantities. The joint demand forecasts for retailers are based on
the same information set denoted by H
t+1
consisting of observed demands up to and including period t from both retailers:
H
t1
= [z
t
; z
t1
; . . .[ = z
1;t
; z
2;t
; z
1;t1
; z
2;t1
; . . .: (14)
The retailers single-period joint forecasts are easily derived if we decompose the future demand into linear combinations of future er-
rors and H
t
. We substitute the joint demand recursively backward in time k times on future demand, z
t+k
, to obtain
z
tk
= q
k1
z
t1

k
j=0
q
j
e
tkj
: (15)
The above additive decomposition yields the following k-period-ahead joint demand forecasts and their associated errors:
^
z
tk
= E(z
tk
[H
t
) = q
k1
z
t1
;
^
e
tk
= z
tk

^
z
tk
=

k
j=0
q
j
e
tkj
:
(16)
It is well-known that the above conditional expectations of future demand given observed demand history form the MMSE forecasts
(Hamilton, 1994). These single-period forecasting results are similar to those for an AR(1) process with matrices replacing scalar
parameters.
Let

D
t
= [

D
1
[t; t L[;

D
2
[t; t L[[
T
represent the vector of lead time demand forecasts and e
D
t
= [D
1
[t; t L[

D
1
[t; t L[;

D
2
[t; t L[

D
2
[t; t L[[
T
be their associated forecasting errors. Dene
U
k
=

k
j=0
q
j
(17)
as the partial sum of matrix sequence q
j
for j P0.
Lemma 1. The retailers lead-time demand forecasts and their associated forecasting errors are

D
t
Ll =

L
k=0
^z
tk
= (U
L1
I)z
t1
; e
D
t
=

L
k=0
U
Lk
e
tk
: (18)
Proof. Summing the k period-ahead forecasts and errors, we have

D
t
Ll =

L
k=0
^z
tk
=

L
k=0
q
k1
_ _
z
t1
; e
D
t
=

L
k=0

k
j=0
q
j
e
tkj
=

L
k=0

Lk
j=0
q
j
_ _
e
tk
: (19)
Because linear operations preserve the MMSE property, the above forecasts correspond to the MMSE forecasts for lead-time demand. h
Corollary 1. The lead-time demand forecasting error covariance matrix can be determined from:
V(e
D
t
) =

L
k=0
U
k
RU
T
k
: (20)
Proof. This is an immediate result of Lemma 1 and the assumption of inter-temporal independence among joint demand errors. h
Clearly, the forecasting error covariance matrix is constant with respect to t. As a result, the safety stock levels we maintain also remain
constant. In matrix notation, the order quantities for period t in Eq. (4) form a vector q
t
= [q
1,t
, q
2,t
]
T
and Eq. (4) can be expressed as
q
t
=

D
t


D
t1
d
t1
: (21)
If we substitute into the above equation the forecasting results in Lemma 1, we obtain
q
t
= U
L1
z
t1
(U
L1
I)z
t2
l: (22)
The order quantities are calculated by linearly combining observed demand in the past two periods. The time-varying order-up-to policy
prescribes a simple linear rule to chase demand in which only the two most recent demand observations matter. From (21), the covariance
matrix for q
t
can be additively decomposed into a sum of the covariance matrix for the demand V(d
t
) and functions of other demand related
parameters. This decomposition is summarized in the following proposition.
Proposition 1. The covariance matrix of retailers order quantity can be calculated by
V(q
t
) = V(d
t
) U
L1
RU
T
L1

L1
i=0
q
i
R(q
i
)
T
: (23)
Proof. The order quantity vector q
t
in (22) can be re-expressed as a linear combination of immediate past error e
t1
and observed demand
two periods ago z
t2
such as
518 X. Zhang, G.J. Burke / European Journal of Operational Research 210 (2011) 514526
q
t
= U
L1
(qz
t2
e
t1
) (U
L1
I)z
t2
l = U
L1
e
t1
U
L1
I U
L1
q ( )z
t2
l = U
L1
e
t1
U
L2
U
L1
( )z
t2
l
= U
L1
e
t1
q
L2
z
t2
l:
Because e
t1
and z
t2
are uncorrelated, we immediately obtain
V(q
t
) = U
L1
RU
T
L1
(q
L2
)V(d
t
)(q
L2
)
T
:
By substituting the YuleWalker equation qV(d
t
)q
T
= V(d
t
) R into the above expression repeatedly, we obtain the stated result.
Note that in the above decomposition formula, terms positioned after V(d
t
) are responsible for any BWE. The general formulas here ap-
ply for arbitrary lead-times and involve complex matrix terms. For ease of exposition, however, additional analytical insights can be
gleaned by examining our model with zero lead-times. Later in this section these insights are extended for general lead-times.
Corollary 2. With zero lead-times, the order quantity covariance matrix is simplied into
V(q
t
) = V(d
t
) qR Rq
T
(24)
from which we obtain the benchmark separate and pooled BWE measures given by
M
1
= 2(q
11
r
11
q
12
r
12
); M
2
= 2(q
22
r
22
q
21
r
12
);
M = M
1
M
2
2(q
11
q
22
)r
12
2(q
12
r
22
q
21
r
11
):
(25)
Proof. When L = 0, we have U
L+1
= I + q. Substituting this into (23), we have
V(q
t
) = V d
t
) (I q)R(I q)
T
qV(d
t
_ _
q
T
R = V(d
t
) qR Rq
T
: (26)
The above expression indicates that the diagonal terms in the following matrix
qR Rq
T
=
2(q
11
r
11
q
12
r
12
) (q
11
q
22
)r
12
q
12
r
22
q
21
r
11
q
11
q
22
( )r
12
q
12
r
22
q
21
r
11
2 q
22
r
22
q
21
r
12
( )
_ _
determine each separate BWE measure, while the off-diagonal elements determine the synergy factor found in the aggregate BWE measure.
Corollary 2 clearly indicates that the impact of demand interaction and cross correlation on the BWE is additive.
More insights can be gleaned from the explicit analytical results in Corollary 2 if we incrementally introduce demand interaction coef-
cients q
12
and q
21
, autocorrelation coefcients q
11
and q
22
, and cross covariance r
12
as three separate causes for the BWE. Six compound-
ing scenarios are examined: three correspond to the separate causes introduced one at a time and three correspond to dual causes with a
combination of two separate causes introduced simultaneously. Table 1 tabulates the parameter settings for these scenarios, their associ-
ated BWE measures, and the resulting pooling factor.
When only a single cause is present, we observe that autocorrelation alone causes a separate BWE measured by M
i
= 2q
ii
r
ii
ifq
ii
> 0. Here,
the pooling factor is zero and no synergy effect exists. Because the joint demand becomes two independent AR(1) processes in this scenario,
the separate BWE measure becomes identical to the relative measure in Eq. (3.5) of Lee et al. (1997a). Interaction alone does not cause a
separate BWE, but it does introduce a synergistic effect if (q
12
r
22
+ q
21
r
11
)< 0. Covariance alone causes neither a separate BWE nor a pool-
ing effect.
For the rst of the three dual-cause scenarios, compounding autocorrelation with covariance does not affect the separate BWE mea-
sures. They remain equal to M
i
= 2q
ii
r
ii
, but a pooling factor is now present. In this case, we have a positive separate BWE when q
ii
> 0.
In addition, if r
12
< 0, the pooling factor 2r
12
(q
11
+ q
22
) is negative, representing a synergy or variability smoothing effect. If we further
impose the condition that r
12
< q
ii
r
ii
=(q
11
q
22
)\i = 1; 2, the synergy factor dominates the sum of the two positive separate BWEs,
resulting in a negative aggregate BWE. Hence, the BWE is present in each order stream but not in the aggregate. The opposite, where
the BWE is present in the aggregate but not in the individual orders, can also occur if q
ii
< 0 and r
12
> q
ii
r
ii
=(q
11
q
22
)\i = 1; 2. In
the second of the three dual-cause scenarios, we compound autocorrelation with demand interaction. Comparing with the previous case,
a different pooling factor 2(q
12
r
22
+ q
21
r
11
) is introduced, while the separate BWE measures remain the same. Similarly when q
ii
> 0 and
q
ij
< (q
ji
q
jj
)r
jj
"i j, the synergistic effect can dominate the sum of the separate BWE, causing a negative aggregate BWE. In the third of
the three dual-cause scenarios, demand interaction and covariance are compounded. Separate bullwhip measures and pooling factors are
both present. Under the conditions of r
12
< 0, andq
21
< 0, the separate BWE measures are positive but aggregate BWE is negative.
These incremental scenarios under zero lead-time suggest that compounding demand correlation or interaction alone with demand
autocorrelation does not impact separate BWE measures. When the separate BWE is positive, we nd that under certain conditions the
Table 1
Compounding cases for benchmark with zero leadtime (i, j = 1, 2, i j).
Compounding scenarios Demand parameter settings Separate BWE measures Pooling factor
Single cause
Autocorrelations q
ii
0, q
ij
= r
12
= 0 M
i
= 2q
ii
r
ii
Zero
Covariance r
12
0, q
ii
= q
ij
= 0 M
i
= 0 Zero
Interactions q
ij
0, q
ii
= 0, r
12
= 0 M
i
= 0 2(q
12
r
22
+ q
21
r
11
)
Dual causes
Autocorrelations + covariance q
ii
0, r
12
0, q
ij
= 0 M
i
= 2q
ii
r
ii
2r
12
(q
11
+ q
22
)
Autocorrelations + interactions q
ii
0, q
ij
0, r
12
= 0 M
i
= 2q
ii
r
ii
2(q
12
r
22
+ q
21
r
11
)
Covariance + interactions r
12
0, q
ij
0, q
ii
= 0 M
i
= 2q
ij
r
12
2(q
12
r
22
+ q
21
r
11
)
X. Zhang, G.J. Burke / European Journal of Operational Research 210 (2011) 514526 519
synergy effect can dominate the sum of separate BWEs and cause a negative aggregate BWE. This observation can be informative for supply
chain managers regarding decisions of individual versus joint replenishment of demand markets.
Even though Corollary 2 produces useful analytical insights for the case of zero lead-times, the exact analytical formula in Eq. (23) be-
comes complex for general lead-times. However, for general lead-times, we extend the conclusions corresponding to scenarios 1 and 2
when compound causes are introduced incrementally.
Proposition 2. For general lead-times,
(a) when there is only demand autocorrelation and demand interaction and correlation are absent, i.e., q
ii
0 and q
12
= q
21
= r
12
= 0, the
pooling factor is zero, no synergy effect exists, and separate BWEs reduce to those in Lee et al. (1997a) where
V(q
i;t
) = V(d
i;t
)
2q
ii
(1 q
L1
ii
) 1 q
L2
ii
_ _
(1 q
ii
)(1 q
2
ii
)
r
ii
: (27)
(b) when only demand correlation is introduced incrementally to a), i.e., q
12
= q
21
= 0, q
ii
0, and r
12
0, the pooling factor becomes non-
zero, and a synergy effect exists for the pooled order quantity if r
12
< 0. Separate BWEs, however, remain the same as in (a).
(c) when only demand interactions are introduced incrementally to (a), i.e., q
12
0, q
21
0, and q
ii
0, but r
12
= 0, the pooling factor is
nonzero and the separate BWE for each retailer can be decomposed additively into that in (a) and another component determined by the
nonzero demand parameters.
Proof. See Online supplement. h
Proposition 2 incrementally introduces several compounding causes one at a time, and suggests that introducing demand interaction
has repercussions on the order stability of each order stream as well as the stability of aggregated orders. Part (c) suggests the possibility
that the existence of separate BWEs does not automatically lead to an aggregate BWE and vice versa, as we found from the special case of
zero leadtime. The existence of a synergy effect with general lead-times will be explored further in numerical analysis Section 5.
4.2. The case of full demand process incorporating price changes
In this section, we consider an additional BWE cause, volatility in product market prices. We use a tilde, ~, to differentiate the quan-
tities in this new setting from those in the benchmark case. The key to our analytical development is stated in the following proposition
containing our nding on the joint time series structure of price and demand streams as a VAR(1) process.
Proposition 3. Let ~z
t
= [z
1;t
; z
2;t
; z
p;t
[
T
. As a vector time series, ~z
t
possesses a VAR(1) structure
~
z
t
= P
~
z
t1
n
t
; n
t
= Av
t
;
P =
q
11
q
12
kb
1p
q
21
q
22
kb
2p
0 0 k
_

_
_

_; A =
1 0 b
1p
0 1 b
2p
0 0 1
_

_
_

_;
E n
t
n
T
t
/
_ _
= 0; \t t
/
;
E n
t
n
T
t
_ _
= X = AVA
T
:
(28)
Proof. See Online supplement. h
We can again resort to YuleWalker equations to nd the unconditional variance of demand and price jointly from
vec[V(
~
z
t
)[ = (I P P)
1
vec[V(n
t
)[: (29)
With the VAR(1) representation, the joint demand and price forecasting can be obtained in exactly the same manner as the benchmark case
by simply making the following substitutions in the baseline forecasting results
~
z
t
= z
t
; P =q; n
tkj
= e
tkj
: (30)
The lead time demand forecasts for our current case can be calculated from

D
i
[t; t L[ Ll
1

D
2
[t; t L[ Ll
2

D
p
[t; t L[ Ll
p
_

_
_

_
=

L
k=0
~
z
tk
=

L
k=0
P
k1
_ _
~
z
t1
; (31)
where

D
p
[t; t L[ is the MMSE forecast of cumulative sum of price during lead time. We can dene

U
k
=

k
j=0
P
j
; H =
1 0 0
0 1 0
_ _
: (32)
The order quantities are found from
q
t
= H

U
L1
~
z
t1
(

U
L1
I)
~
z
t2

~
l
_ _
= H

U
L1
n
t1


U
L2


U
L1
_ _
~
z
t2

~
l
_ _
; (33)
520 X. Zhang, G.J. Burke / European Journal of Operational Research 210 (2011) 514526
where ~ l = [l
1
; l
2
; l
p
[
T
. The covariance matrix for the bracketed term can be calculated from
V(q
t
) = V(d
t
) H

U
L1
X

U
T
L1

L1
i=0
P
i
X(P
i
)
T
_ _
H
T
: (34)
Again, we examine the special case of zero lead times for analytical insights. In this case, the covariance matrix for order quantities in Eq. (34)
simplies to
V(q
t
) = V(d
t
) H[PXXP
T
[H
T
:
Clearly, when b
1p
= b
2p
= r
1p
= r
2p
= 0, the price-sensitive case reduces to the benchmark case. The following proposition shows that we can
isolate the compounding effect of price sensitivity on the BWE from the benchmark case.
Proposition 4. Let r
p
= [r
1p
, r
2p
, r
pp
]
T
. With zero lead times, the price-sensitive separate and aggregate BWE measures are given by
M
1
= 2(q
11
r
11
q
12
r
12
) 2p
T
1
r
p
; M
2
= 2(q
22
r
22
q
21
r
12
) 2p
T
2
r
p
;
M = M
1
M
2
2[(q
11
q
22
)r
12
q
12
r
22
q
21
r
11
[ 2p
T
3
r
p
;
where
p
1
=
(2q
11
k)b
1p
q
12
b
2p
q
12
b
1p
(q
11
k)b
2
1p
q
12
b
1p
b
2p
_

_
_

_
; p
2
=
q
21
b
1p
(2q
22
k)b
2p
q
21
b
1p
(q
22
k)b
2
2p
q
21
b
1p
b
2p
_

_
_

_
; p
3
=
2q
21
b
1p
(q
11
q
22
k)b
2p
2q
12
b
2p
(q
11
q
22
k)b
1p
q
21
b
2
1p
q
12
b
2
2p
(q
11
q
22
2k)b
1p
b
2p
_

_
_

_
:
Proof. See Online supplement. h
Comparing to the benchmark case with zero lead-time, the compounding impact on the bullwhip from price changes is additive with its
incremental effects quantied by the terms 2p
T
1
r
p
; 2p
T
2
r
p
, and 2p
T
3
r
p
. To gain insight into how these terms relate to demand parameters
associated with price dynamics, we compound price changes with single-cause and dual-cause scenarios examined for the benchmark
in Table 1. Table 2 summarizes analytical results obtained from setting the appropriate demand parameters to a zero value. A quick com-
parison between Table 1 and 2 indicates that price changes have a complex impact on the separate BWE measures as well as the pooling
factor. The case involving just price changes as a BWE cause sheds light on this additional layer of complexity. In this even more special
setting, we eliminate all BWE causes from the benchmark by setting q
11
= q
22
= q
12
= q
21
= r
12
= 0 and keeping b
ip
> 0 and k > 0. We then
have
M
i
= 2kb
ip
(b
ip
r
pp
r
ip
);
M = M
1
M
2
2k(2b
1p
b
2p
r
pp
b
1p
r
2p
b
2p
r
1p
) = M
1
M
2
2k[b
1p
(b
2p
r
pp
r
2p
) b
2p
(b
1p
r
pp
r
1p
)[:
Table 2
Compounding price changes with benchmark scenarios with zero leadtime (i, j = 1, 2, i j).
Benchmark compounding
scenarios
Demand
parameter
settings
Separate BWE measures Pooling factor
Benchmark single cause + price effect
Autocorrelation q
ii
0
b
ip
0; r
ip
0
q
ij
= r
12
= 0
M
i
= 2q
ii
r
ii
2(2q
ii
k)b
ip
r
ip
2(q
ii
k)b
2
ip
rpp
2(q
11
q
22
k)b
2p
r
1p
2(q
11
q
22
k)b
1p
r
2p
2(q
11
q
22
2k)b
1p
b
2p
rpp
Covariance r
12
0
b
ip
0; r
ip
0
q
ii
= q
ij
= 0
M
i
= 2kb
2
ip
rpp 2kb
ip
r
ip
2kb
2p
r
1p
2kb
1p
r
2p
4kb
1p
b
2p
rpp
Interaction q
ij
0
b
ip
0; r
ip
0
q
ii
= 0; r
12
= 0
M
i
= 2(kb
ip
q
ij
b
jp
)r
ip
2q
ij
b
ip
r
jp
2(kb
2
ip
q
ij
b
ip
b
ip
)rpp
2(q
12
r
22
q
21
r
11
) 2(2q
21
b
1p
kb
2p
)r
1p
2(2q
12
b
2p
kb
1p
)r
1p
2(q
21
b
2
1p
q
12
b
2
2p
2kb
1p
b
2p
)rpp
Benchmark dual causes + price effect
Autocorrelation + covariance q
ii
0; r
12
0
b
ip
0; r
ip
0
q
ij
= 0
M
i
= 2q
ii
r
ii
2(2q
ii
k)b
ip
r
ip
2(q
ii
k)b
2
ip
rpp
2(q
11
q
22
)r
12
2(q
11
q
22
k)b
2p
r
1p
2(q
11
q
22
k)b
1p
r
2p
2(q
11
q
22
2k)b
1p
b
2p
rpp
Autocorrelation + interactions q
ii
0; q
ij
0
b
ip
0; r
ip
0
r
12
= 0
M
i
= 2q
ii
r
ii
2p
T
1
rp 2(q
12
r
22
q
21
r
11
) 2p
T
3
rp
Covariance + interactions r
12
0; q
ij
0
b
ip
0; r
ip
0
q
ii
= 0
M
i
= 2q
ij
r
12
2q
ij
b
ip
r
jp
2(kb
ip
q
ij
b
jp
)r
ip
2(kb
2
ip
q
ij
b
ip
b
ip
)rpp
2(q
12
r
22
q
21
r
11
) 2(2q
21
b
1p
kb
2p
)r
1p
2(2q
12
b
2p
kb
1p
)r
1p
2(q
21
b
2
1p
q
12
b
2
2p
2kb
1p
b
2p
)rpp
X. Zhang, G.J. Burke / European Journal of Operational Research 210 (2011) 514526 521
The separate BWE depends on the sign of the price autocorrelation coefcient k and the parenthesized term (b
ip
r
pp
r
ip
). We experience a
separate BWE under the following cases:
1. k is positive and r
ip
is negative.
2. k is positive and r
ip
is positive, but b
ip
r
pp
> r
ip
.
3. k is negative and r
ip
is positive, but b
ip
r
pp
< r
ip
.
The key observation is that the pooling factor and the separate BWE work in tandem and reinforce each other, namely, when both sep-
arate BWEs are positive (negative), the pooling factor is also positive (negative). Hence, price changes can either enhance or dampen the
pooling factor and BWEs beyond the benchmark case. The intuition behind this special case result is that price changes induce an indirect
demand autocorrelation, which is at the heart of our observation here. To see this, the joint demand system in (8) reduces to
z
1;t
= b
1p
z
p;t
e
1;t
; z
2;t
= b
2p
z
p;t
e
2;t
;
from which we can derive "i = 1,2
z
i;t
kz
i;t1
= b
ip
(z
p;t
kz
p;t1
) e
i;t
ke
i;t1
= b
ip
e
p;t
e
i;t
ke
i;t1
: (35)
Hence, each demand stream indirectly possesses an ARMA(1, 1) structure. This induced demand autocorrelation dynamic drives the separate
BWE we obtained here. Because both demand streams share a common price error term, the two ARMA(1, 1) processes are not uncorrelated
even if the joint demand errors are uncorrelated, and this cross-correlation induced by the price error is behind the pooling factor we derived
for this special case.
The conditions for a synergistic effect for the cases in Table 2 become more complicated than those in Table 1. For example, in the rst
case when demand autocorrelation is compounded with price changes, it is easy to establish that when q
ii
> 0, b
ip
r
pp
2r
ip
> 0, and k > 0,
we have M
i
> 0 and the pooling factor is positive, and pooling orders increases the BWE; when these inequalities are reversed, we have the
opposite effect. Hence, there are demand parameter regions in which the pooling factor has the same sign as the separate BWE, exhibiting a
reinforcing impact on M
i
. In addition, if we introduce a non-zero demand covariance incrementally (the fourth case in Table 2), the separate
BWE measures do not remain, but the pooling factor involves an extra additive term 2(q
11
+ q
22
)r
12
.
For general lead-times, the compounding effect of price changes becomes complicated. Fortunately, we can still establish an extension
of Proposition 4 in that the compounding effect from price changes on the BWE and pooling factor are incremental to the benchmark case.
Proposition 5. Under general lead-times, the introduction of price-sensitivity in the joint demand has an additive incremental impact on separate
and aggregate BWE measures beyond the benchmark case.
Proof. See Online supplement. h
The impact of price changes on BWE is complex with arbitrary leadtimes. We explore its shape with a numerical study in the next
section.
5. Numerical analysis of nonzero lead times
We consider symmetric retailers to reduce the number of demand parameters for our numerical study. The demand autocorrelation,
demand interaction, and price sensitivity parameters are assumed identical between the two retailers. Therefore, all sensitivity or compar-
ative results reported here are attributed to simultaneous changes in two parameters that are kept equal. We explore the numerical ranges
of the BWE measures for each of the compounding cases when we vary demand parameters. The sensitivity of our analytical results ob-
tained for zero lead time is evaluated relative to longer lead times and clues to managerial insights are discussed.
5.1. Benchmark
Symmetric retailers impose following parameter restrictions for our numerical analysis
q
11
= q
22
; q
21
= q
12
; r
11
= r
22
;
and for the zero lead time case the sum of separate BWE measures M
1
+ M
2
and the aggregate BWE measure M in (25) become
M
1
M
2
= 4(q
11
r
11
q
12
r
12
);
M = M
1
M
2
4q
11
r
12
4q
12
r
11
= 4(q
11
q
12
)(r
11
r
12
):
(36)
These analytical results indicate the following trends:
1. M is increasing in q
12
because r
11
+ r
12
> 0 for symmetric retailers.
2. M
1
+ M
2
is increasing in q
12
if r
12
> 0.
3. A positive q
11
causes a positive separate BWE while a negative q
11
, and vice versa.
4. The synergy effect exists if 4q
11
r
12
+ 4q
12
r
11
< 0 based upon (36) or
q
12
> d =
q
11
r
12
r
11
:
We call d the synergistic threshold. It is positive (negative) when demand autocorrelation and cross-covariance have opposite (the same)
signs.
522 X. Zhang, G.J. Burke / European Journal of Operational Research 210 (2011) 514526
For lead-times L = 1, 2, . . ., 8, our numerical results show that these analytical trends for the zero leadtime continue to hold except for
trend 2 when a stronger interaction, either positive or negative, increases M
1
+ M
2
for L P2 As an example, Fig. 1 depicts M
1
+ M
2
and
M as a function of demand interaction when there is a one-period lead time (L = 1). For ease of comparison, the four juxtaposed plots cor-
respond to four different parameter settings while we change the signs of demand autocorrelation and interaction with a positive demand
cross-correlation: (1) q
11
= q
22
= 0.5 and r
12
= 0.5; (2) q
11
= q
22
= 0.5 and r
12
= 0.5; (3) q
11
= q
22
= 0.5 and r
12
= 0.5; (4) q
11
= q
22
= 0.5
and r
12
= 0.5. Without loss of generality, we assume r
11
= r
22
= 1. The plots are made by varying q
12
and q
21
simultaneously and main-
taining q
12
= q
21
to reect the assumption of symmetric retailers. For stationarity, q
12
and q
21
can only vary within (0.5, 0.5). Fig. 2 clearly
reveals the same analytical trends for zero lead time reported in Table 1, suggesting a synergistic threshold at which the M curve crosses
the M
1
+ M
2
curve from below.
5.2. Full demand model with price-changes
With each of the four benchmark settings examined in 5.1, we compare the price-sensitive case by varying the signs of price auto-
regression coefcient k and the price-demand cross correlations r
ip
. Again, we investigated the impact of price changes on the general
trends observed for the benchmark case for L = 1, 2, . . . , 8. Fig. 2 summarizes the numerical ndings when L = 1. It contains four rows of plots
and each row corresponds to one of four demand parameter settings as in Fig. 1. Within each row, the four charts represent different
parameter settings for k and r
ip
and serve as comparisons of the price sensitive model to the respective benchmarks. Without loss of gen-
erality, we assume r
pp
= 1 and b
1p
= b
1p
= 0.5. Each column of plots in Fig. 2 can be visually compared with its corresponding plot in Fig. 1 to
isolate the incremental compounding effect of price change beyond the compounding that may occur in benchmark cases. For this purpose,
the plots for each case use the same scale. The following observations can be made:
1. The rst column of plots in Fig. 2 shows that when k and r
ip
are both positive, price changes appear to increase the separate and aggre-
gate BWEs beyond the benchmark case. This incremental effect seems to be more pronounced when q
11
and q
22
are positive and
demand interaction is strong positively.
2. The fourth column of plots shows that when k and r
ip
are both negative, price changes appear to suppress the separate BWE beyond the
benchmark case, especially when q
11
and q
22
are positive and demand interaction is strong positively or when q
11
and q
22
are negative
and demand interaction is strong negatively.
Fig. 1. The benchmark numerical results: L = 1.
X. Zhang, G.J. Burke / European Journal of Operational Research 210 (2011) 514526 523
3. In the remaining two columns, when k and r
ip
have the opposite signs, we have less clear-cut patterns. In the second column, where k is
positive and r
ip
is negative, we observe that price changes exacerbate the BWE except when demand autocorrelation and interaction are
negative. In the third column, where k is negative and r
ip
is positive, the incremental effect from price changes is positive. The incre-
mental effect, however, seems negligible when demand autocorrelations q
11
and q
22
are also positive.
Fig. 2. Compounding impact of price change on bullwhip effect: L = 1.
524 X. Zhang, G.J. Burke / European Journal of Operational Research 210 (2011) 514526
4. The pooling factors continue to exist where the synergistic threshold seems to be decreased (increased) relative to the benchmark case
when k is positive (negative).
For lead times longer than one period, similar patterns were observed. In summary, we nd that k and r
ip
are complementary and rein-
forcing pairs in causing additional instability in the orders. The direction of the impact depends on the signs of demand autocorrelation and
interaction. The complex array of possible patterns illustrated is not surprising, given that inuences of price changes happen through a
deterministic price effect and price demand cross-correlation. This is consistent with analytical results for zero lead-time, which involves
complex relationships between the incremental effect of price changes on the BWE and the underlying demand parameters.
6. Summary and managerial implications
We introduce a setting in this paper where multiple demand streams with price effects may interact both stochastically and determin-
istically. As such, key demand characteristics can have a compounded impact on the BWE. We quantify this jointly occurring BWE and
show that order variability propagated via individual causes can be combined in an additive manner to produce an overall BWE. We spe-
cically illustrate how interrelated dynamics of demand stream diversity and price shocks simultaneously determine the BWE beyond a
single-demand stream case with only demand autocorrelation. Several important managerial insights are revealed. We demonstrate both
analytically for the zero-lead time case and numerically for other lead times that demand interaction and correlation across multiple de-
mand streams can compound to exacerbate or mitigate the BWE. Interestingly, individual demand streams may in isolation increase var-
iability in individual orders, but when combined may mitigate the BWE due to a synergy effect (i.e., M
1
M
2
> M). The converse may also
occur, namely, individual demand streams may each separately not induce a large order variability, but in combination may produce a lar-
ger aggregate BWE (i.e., M
1
M
2
< M).
Introducing price changes into the analysis shows that price autocorrelation and price cross-correlation with demand streams provide
additional checks and balances that can help dampen the BWE for each individual demand stream or for aggregated demand. Our rm level
ndings of cancellation effects in the aggregate provide a possible explanation for the (Cachon et al., 2007) empirical nding that a signif-
icant BWE in aggregated industry level data is undetectable.
Supply chain managers can employ our analytical results to gauge the combined BWE when crafting policies to mitigate its negative
impact on supply chain operations. First, from a marketing perspective, it may be benecial to take a portfolio approach to market selec-
tion, in which we combine demand streams with interaction coefcients less than the synergistic threshold, as illustrated in Figs. 1 and 2.
Demand interactions and cross correlations can profoundly impact the variability of aggregated demand information. Therefore, market
mix with a desired interaction that can dampen the variability of the aggregated order may warrant explicit consideration in supply chain
design. Additionally, in competitive markets the auto-correlative effects of exogenous prices should be considered. These demand factors
may be protably employed to mitigate BWE and supply chain risks in much the same way that investment portfolios are balanced and
rebalanced to gain benets from diversication.
Second, our ndings may call for multi-pronged bullwhip mitigation strategies. Rather than implementing a single dampening tactic
rms may nd it benecial to develop pairs or trios of tactics whose combined effect lessens demand distortion. For example, a rm
may undertake a ght re with re approach to dealing with an identied bullwhip cause by introducing another counteractive cause to
the supply chain. In isolation if each cause creates demand volatility, but together the two causes may enhance demand stability. Con-
versely, the dynamics of compound causes of the BWE may also necessitate incremental mitigation strategy (i.e., combating causes one
at a time) by splitting compounded causes.
Third, in broader operational contexts, globalization and product proliferation require more accurate and timely forecasts. Many orga-
nizations seek to jointly forecast multiple demand streams (geographic regions, product markets, product types) over a dened time hori-
zon. For example, many ERP systems have built-in functionality for handling this complexity and our results will have practical
implications to managing the demand stability in such systems. Furthermore, our model quanties the benets of demand stream diver-
sication and risks of compounded BWE causes. As such, our results can guide decision-making in regard to several operational areas
including supply chain structure, lead-time policies, product bundling and assignment of replenishment depots to satisfy downstream
markets.
Areas for future work can focus on either empirical validation of our analytical results or theoretical extension of our model. In the
empirical front, rm-level demand and order data can be collected to estimate key parameters of our model, which can be used to validate
our ndings on the BWE. Similar validation can also be performed using industry-level data as done in Cachon et al. (2007), assuming that
our model is applicable to aggregate industries and a macro-economic environment. Analytically, our work can extended to incorporate
order batching and gaming rations as additional compound causes of the BWE. In addition, the research community can use our model
to explore the links between BWE and operational hedging research areas. One such link may be to further investigate how BWE may
be inuenced by the hedging possibilities inherent in multiple demand streams and multiple compound BWE causes.
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in the online version, at doi:10.1016/j.ejor.2010.09.030.
References
Aviv, Y., 2001. The effect of collaborative forecasting on supply chain performance. Management Science 47, 13261343.
Aviv, Y., 2003. A time-series framework for supply-chain inventory management. Operations Research 51 (2), 210227.
Barrett, L., 2001. Ciscos $2.25 billion mea culpa. <http://www.news.com/Ciscos-2.25-billion-mea-culpa/2100-1033_3-257278.html> Retrieved 23.12.07.
Boyabatl, O., Toktay, L.B., 2004. Operational Hedging: A Review with Discussion, Technical Report. INSEAD, Fontainebleau, France.
Burke, G.J., Carrillo, J.E., Vakharia, A.J., 2007. Single vs. multiple supplier sourcing strategies. European Journal of Operational Research 182 (1), 95112.
Cachon, G., Randall, T., Schmidt, G., 2007. In search of the bullwhip effect. Manufacturing & Services Operations Management 9 (4), 457479.
Chen, Frank, Drezner, Zvi, Ryan, Jennifer K., Simchi-Levi, David, 2000. Quantifying the bullwhip effect in a simple supply chain. Management Science 46 (3), 436443.
X. Zhang, G.J. Burke / European Journal of Operational Research 210 (2011) 514526 525
Corbett, C.J., Rajaram, K., 2006. A generalization of the inventory pooling effect to nonnormal dependent demand. Manufacturing & Services Operations Management 8 (4),
351358.
Dejonckheere, J., Disney, S.M., Lambrecht, M.R., Towill, D.R., 2003. Measuring and avoiding the bullwhip effect: A control theoretic approach. European Journal of Operational
Research 147 (3), 567590.
Dejonckheere, J., Disney, S.M., Lambrecht, M.R., Towill, D.R., 2004. The impact of information enrichment on the bullwhip effect in supply chains: A control engineering
perspective. European Journal of Operational Research 153 (3), 727750.
Dekimpe, M.G., Hanssens, D.M., 1995. Empirical generalizations about market evolution and stationarity. Marketing Science 14 (3), G10921.
Dekimpe, M.G., Hanssens, D.M., Silva-Risso, J.M., 1998. Long-run effect of price promotions in scanner markets. Journal of Econometrics 89 (12), 269291.
Disney, S.M., Lambrecht, M.R., 2007. On replenishment rules, forecasting, and the bullwhip effect in supply chains. Foundations and Trends

in Technology, Information and


Operations Management 2 (1), 180.
Federgruen, A., Heching, A., 1999. Combined pricing and inventory control under uncertainty. Operations Research 47 (3), 454475.
Forrester, J.W., 1958. Industrial dynamics A major breakthrough for decision making. Harvard Business Review 36 (4), 3766.
Forrester, J.W., 1961. Industrial Dynamics. MIT Press, Cambridge, MA.
Gaur, V., Giloni, A., Seshadri, S., 2005. Information sharing in a supply chain under ARMA demand. Management Science 51 (6), 961969.
Gilbert, K., 2005. An ARIMA supply chain model. Management Science 51 (2), 305310.
Graves, S.C., 1999. A single-item inventory model for a non-stationary demand process. Manufacturing & Services Operations Management 1 (1), 5061.
Hamilton, James D., 1994. Time Series Analysis. Princeton University Press, Princeton, NJ.
Hanssens, D.M., Parsons, L.J., Schultz, R.L., 2001. Market Response Models: Econometric and Time Series Analysis, second ed. Kluwer Academic Publishers, Boston, MA.
Harrison, J.M., Van Mieghem, J.A., 1999. Multi-resource investment strategies: Operational hedging under demand uncertainty. European Journal of Operations Research 113
(1), 1729.
Johnson, G.D., Thompson, H.E., 1975. Optimality of myopic inventory policies for certain dependent demand processes. Management Science 21 (11), 13031307.
Jaksic, M., Rusjan, B., 2008. The effect of replenishment policies on the bullwhip effect: A transfer function approach. European Journal of Operations Research 184, 946961.
Kouvelis, P., Chambers, C., Wang, H., 2006. Supply chain management research and production and operations management: Review, trends, and opportunities. Production
and Operations Management 15 (3), 449469.
Lee, H.L., Padmanabhan, P., Whang, S., 1997a. Information distortion in a supply chain: The bullwhip effect. Management Science 43, 546558.
Lee, H.L., Padmanabhan, P., Whang, S., 1997b. Bullwhip effect in a supply chain. Sloan Management Review 38 (Spring), 93102.
Lee, H.L., Padmanabhan, P., Whang, S., 2004. Comments on information distortion in a supply chain: The bullwhip effect. Management Science 50 (12), 18871893.
Lee, H.L., So, K.C., Tang, C.S., 2000. The value of information sharing in a two-level supply chain. Management Science 46, 626643.
Ltkepohl, H., 2005. New Introduction to Multiple Time Series Analysis. Springer, Berlin.
Ouyang, Y., Daganzo, C.F., 2006. Characterization of the bullwhip effect in linear, time-invariant supply chains: Some formulae and tests. Management Science 52 (10), 1544
1556.
Ouyang, Y., 2007. The effect of information sharing on supply chain stability and the bullwhip effect. European Journal of Operational Research 182 (3), 11071121.
Ouyang, Y., Li, X., 2010. The bullwhip effect in supply chain networks. European Journal of Operational Research 201, 799810.
Ozelkan, E.C., Cakanyldrm, M., 2009. Reverse bullwhip effect in pricing. European Journal of Operational Research 192, 302312.
Raghunathan, S., 2001. Information sharing in a supply chain: A note on its value when demand is non-stationary. Management Science 47, 605610.
Schwartz, E., Smith, J.E., 2000. Short-term variations and long-term dynamics in commodity prices. Management Science 46 (7), 893911.
Sterman, J.D., 1989. Modeling managerial behavior: Misperceptions of feedback in a dynamic decision making experiment. Management Science 35 (3), 321339.
Tsay, R.S., 2002. Analysis of Financial Time Series. John Wiley & Sons, New York, NY.
Wu, Y., Katok, Elena., 2006. Learning, communication, and the bullwhip effect. Journal of Operations Management 24 (6), 839850.
Mieghem, Van, 2007. Risk mitigation in newsvendor networks. Management Science 53 (8), 12691288.
Veinott, A., 1965. Optimal policy for a multi-product dynamic nonstationary inventory problem. Management Science 12, 206222.
Zhang, X., Zhao, Y., 2010. The impact of external demand information on parallel supply chains with interacting demand. Production and Operations Management 19 (4), 463
479.
Zhang, X., 2004a. The impact of forecasting methods on the bullwhip effect. International Journal of Production Economics 88 (1), 1527.
Zhang, X., 2004b. Evolution of ARMA demand in supply chains. Manufacturing and Services Operations Management 6 (2), 195198.
Zipkin, P.H., 2000. Foundations of Inventory Management. McGraw-Hill, New York, NY.
526 X. Zhang, G.J. Burke / European Journal of Operational Research 210 (2011) 514526

Potrebbero piacerti anche