Sunteți pe pagina 1din 12

European Journal of Operational Research 201 (2010) 799810

Contents lists available at ScienceDirect

European Journal of Operational Research


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

Stochastics and Statistics

The bullwhip effect in supply chain networks


Yanfeng Ouyang *, Xiaopeng Li
Department of Civil and Environmental Engineering, University of Illinois at Urbana-Champaign, Urbana, IL 61801, United States

a r t i c l e i n f o a b s t r a c t

Article history: This paper analyzes the propagation and amplication of order uctuations (i.e., the bullwhip effect) in
Received 3 June 2007 supply chain networks operated with linear and time-invariant inventory management policies. The sup-
Accepted 31 March 2009 ply chain network is allowed to include multiple customers (e.g., markets), any network structure, with
Available online 8 April 2009
or without sharing information. The paper characterizes the stream of orders placed by any supplier for
any stationary customer demand processes, and gives exact formulas for the variance of the orders placed
Keywords: and the amplication of order uctuations. The paper also derives robust analytical conditions, based
Supply chain management
only on inventory management policies, to predict the presence of the bullwhip effect for any network
Network
The bullwhip effect
structure, any inventory replenishment policies, and arbitrary customer demand processes. Numerical
Frequency domain analysis examples show that the analytical results accurately quantify the bullwhip effect; managerial insights
are drawn from the analysis. The methodology presented in this paper generalizes those in previous stud-
ies for serial supply chains.
2009 Elsevier B.V. All rights reserved.

1. Introduction

Extensive research has been conducted to analyze the bullwhip effect in supply chains (sometimes called the instability problem), a
phenomenon in which the uctuations in the order sequence grow larger for suppliers farther away from the customers. Observations in
industry operations (Magee, 1956; Forrester, 1958; Forrester, 1961; Magee and Boodman, 1967), macroeconomic data (Holt et al., 1960;
Blinder, 1986; Kahn, 1987; Ramey, 1991; Naish, 1994; Baganha and Cohen, 1998), and simulated experiments (such as the Beer Game
(Sterman, 1989; Goodwin and Franklin, 1994; Kaminsky and Simchi-Levi, 1998)) reveal huge extra supply chain costs due to this problem
(Cooke, 1993; Lee et al., 1997a). The majority of existing research uses a statistical approach to derive the variance of the orders placed by
the supplier under certain customer demand process. Interesting managerial insights are obtained by identifying various operational
causes and quantifying how various factors affect the variance amplication (Lee et al., 1997b; Baganha and Cohen, 1998; Graves,
1999; Chen et al., 2000a,b; Zhang, 2004; Gaur et al., 2005; Gilbert, 2005).
A system control framework was recently introduced to study the bullwhip effect in the frequency domain (Daganzo, 2001, 2003,
2004; Dejonckheere et al., 2003, 2004; Ouyang, 2005). Variance formulas for the orders by any supplier have been derived for multi-stage
serial chains and any ergodic customer demand (Ouyang and Daganzo, 2006a, 2008). When customer demand is hard to predict or con-
trol, robust diagnostic tests (i.e., tests that hold independently of the customer demand) are presented for the existence of the bullwhip
effect at any stage of the chain (Daganzo, 2001, 2003, 2004; Ouyang and Daganzo, 2006a). It is demonstrated in (Ouyang and Daganzo,
2006a) that the bullwhip effect (i.e., variance amplication) may be missing in the rst a few stages (e.g., retailers) then arise in other
stages (e.g., wholesalers), even if the same operating policies are used. We suspect that this is consistent with the recent empirical obser-
vations at the industry level (Cachon et al., 2007). Robust stability analysis is shown to be appealing, especially for multi-echelon chains,
because operating strategies based on these results ensure supply chain stability no matter what the customer does. Similar robust results
are also found for chains with stochastic supplier behavior and operating uncertainties (Boccadoro et al., 2006; Ouyang and Daganzo,
2008). Various strategies to avoid or mitigate the bullwhip have been discussed (Disney et al., 2006; Ouyang and Daganzo, 2006b; Ouyang
et al., 2006).
Previous analyses generally focus on serial supply chains. Most empirical systems, however, have a network topology in which each
suppliers ordering decisions may be inuenced directly by orders from multiple neighbors, or indirectly via network-wide information
sharing. More importantly, demands may arise from multiple customers (or markets) at different locations of the supply network, and the

* Corresponding author.
E-mail address: yfouyang@illinois.edu (Y. Ouyang).

0377-2217/$ - see front matter 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2009.03.051
800 Y. Ouyang, X. Li / European Journal of Operational Research 201 (2010) 799810

Fig. 1. A general supply network.

interactions (e.g., potential competitions or collaborations) among suppliers and customers should be addressed. Serial supply chains can-
not accommodate these issues.
This paper aims to analyze the bullwhip effect for supply chains with general network topology, general linear ordering policies (infor-
mation sharing scheme), and various customer demand(s). General network topology will enable us to model complex interactions among
suppliers, and to address possible correlations among multiple customer demands. Robust formulas to test the existence of the bullwhip
effect, which do not require knowledge of the demand process, will be presented. These formulas will enable us to device robust ordering
policies or operations that will avoid the bullwhip effect no matter what the customers do. This paper will also present formulas to char-
acterize any order streams in the supply network when customer demand processes are known and stationary (even non-ergodic) this
signicantly strengthens ndings in previous studies (e.g., Ouyang and Daganzo, 2006a). The results include exact variance formulas for all
upstream supplier orders and a formula for the bullwhip effect magnitude.
The exposition of the paper is as follows: Section 2 formulates the problem, introduces notation, and proposes metrics for robust anal-
ysis in different scenarios; Section 3 presents analytical derivations and formulas; Section 4 demonstrates numerical examples on applying
the framework to network systems; nally, Section 5 discusses possible extensions.

2. Formulation

2.1. System dynamics

Fig. 1 depicts a general supply network G N; E, where N : NS [ ND [ NV includes a set of intermediate suppliers, NS , serving a
set of customers, ND , from a set of vendors NV . Put a directed arc i; k 2 E; k i, in the graph if partner i 2 NS [ ND orders from
k 2 NS [ NV . Without losing generality, we assume that NS ; ND and NV are disjoint; i.e., the nodes in ND (customers) have zero
in-degrees, those in NV have zero out-degrees, and those in NS have nonzero in- and out-degrees. There are four disjoint subsets of arcs,

E1 : fi; k 2 E : i 2 ND ; k 2 NV g;
E2 : fi; k 2 E : i 2 ND ; k 2 NS g;
E3 : fi; k 2 E : i 2 NS ; k 2 NS g;
E4 : fi; k 2 E : i 2 NS ; k 2 NV g:

For a generic supplier i 2 NS , its inventory position xi t (including in-transit inventory) and in-hand inventory yi t satisfy conservations
according to orders placed and received. If i; k 2 E, supplier i orders uik t items from k at discrete times t . . . ; 2; 1; 0; 1; 2; . . ., and re-
ceives the items after a constant lead time, lik 0; 1; 2; . . . (i.e., assuming that the upstream always have ample in-stock items, as in (Gavirn-
eni et al., 1999; Lee et al., 2000; Chen et al., 2000a,b; Ouyang and Daganzo, 2006a; Ouyang, 2007)). The conservation equations for the
suppliers inventory position, xi t, and for the in-stock inventory, yi t, are:
X X
xi t 1 xi t uis t  uri t; 8i 2 NS ; 1
s:i;s2E r:r;i2E
X X
yi t 1 yi t uis t  lis  uri t; 8i 2 NS : 2
s:i;s2E r:r;i2E

Eqs. (1) and (2) dene the system dynamics for the supply chain when complemented with the suppliers reorder policies; i.e., the rec-
ipes for determining uik t; i; k 2 E3 [ E4 , from available information.
At time t, the complete information set for the entire network includes the inventory records xi , yi , for all i 2 NS up to period t, and
orders urs ; 8r; s 2 E, up to period t  1:
" # " #
[ [
It : fxi t; xi t  1; . . . ; xi 1; yi t; yi t  1; . . . ; yi 1g [ furs t  1; urs t  2; . . . ; urs 1g :
i2NS r;s2E
Y. Ouyang, X. Li / European Journal of Operational Research 201 (2010) 799810 801

If information is shared across the network, every supplier may determine its order quantities based on any subset of It.1
As in (Ouyang and Daganzo, 2006a; Ouyang, 2007), we focus on a broad family of linear and time-invariant (LTI) policies that are based
on It; i.e., uik t is a time-independent linear function of the elements in It. We also assume that such policies are proper; i.e., when the
sizes of the orders received are constant over time, the supplier inventories tend to equal equilibrium values that are independent of the
initial conditions.2 The unit shift operator P satises Pm xi t xi t  m; 8t and 8m 0; 1; . . .. The most general LTI expression of policy for
uik ; i; k 2 E3 [ E4 is:
X  r  X  rs 
uik t cik Aik Pxr t Brik Pyr t C ik Purs t  1 ; 3
r2NS r;s2E

where cik is a real number; and Arik P; Brik P and C rs r r rs


ik P are polynomials with real coefcients. Polynomials Aik P, Bik P and C ik P, respec-
tively indicate how uik is determined based on supplier rs inventory history, xr ; yr , and its past orders urs . A general denition of
Arik P; Brik P and C rs
ik P can represent any possible LTI ordering policies that may utilize any shared (or local) information. For example, an
order-up-to policy for uik based on supplier is (decentralized) local information, where the up-to level is forecasted as the moving-average
P
of the total demand r:r;i2E uri received over the last m periods, is denoted by (3) with
 (
1; if r i; lik
1 P    P m1 ; if s i;
Arik P ; Brik P 0; 8 r; C rs
ik P m
0; otherwise; 0; otherwise:

System dynamics (1)(3) implicitly allow for negative orders or negative inventories, which may not be realistic in certain businesses. It
has been demonstrated, however, that such implicit assumptions do not signicantly affect the analysis of order stability (Chen et al.,
2000b; Daganzo, 2001, 2003), especially when the focus is on the amplication of small uctuations. The relaxation of these assumptions
will be left for future studies.

2.2. Bullwhip effect metrics

The assumption of properness ensures that if every supplier of a network, e.g., i 2 NS , uses proper policies (as assumed) and if it re-
ceives orders of constant sizes uri t u1 ri ; 8r; i 2 E, at every time period t, then a steady state (or equilibrium) must arise where supplier
i places orders of constant sizes, uis t u1 1
is ; 8i; s 2 E at all t, while maintaining steady inventory positions xi and in-stock inventories yi .
1

The equilibrium inventories, x1 i and y 1


i ; 8i 2 N S , are related to the equilibrium order sizes, u1
rs ; 8r; s 2 E via the system dynamics (1)
(3); i.e.:
X X
0 u1
is  u1
ri ; 8i 2 NS ; 4
s:i;s2E r:r;i2E
X  r r
 X rs
u1
ik cik Aik 1x1 1
r Bik 1yr C ik 1u1
rs ; 8i; k 2 E3 [ E4 : 5
r2NS r;s2E

The inventories and orders can be expressed in terms of deviations (errors) from their values at an equilibrium,
xi t : xi t  x1
 1 1
i ; yi t : yi t  yi ; 8i 2 NS , and urs t : urs t  urs ; 8r; s 2 E. Then, by subtracting (5) from (3) and (4) from (1) and
 
(2), the system dynamics are equivalently represented by the following equations:
X   X  rs 
 ik t
u Arik Pxr t Brik Py
r t  rs t  1 ;
C ik Pu i; k 2 E3 [ E4 ; 6
r2NS r;s2E
X X
xi t 1 xi t  is t 
u  ri t;
u 8i 2 NS ; 7
s:i;s2E r:r;i2E
X X
y i t
i t 1 y  is t  lis 
u  ri t;
u 8i 2 NS : 8
s:i;s2E r:r;i2E

We assume without losing generality that the system is in this equilibrium (i.e., xi t yi t u
 rs t 0) for all t 1; . . . ; 0. Any realization
of the customer demands fu  rs t : r; s 2 E1 [ E2 g1
t0 denes a unique set of upstream order sequences  ik t : i; k 2 E3 [ E4 g1
fu t0 from (6)
(8).
When customer demands are known, all order sequences in the network can be determined from the dynamics equations. The bullwhip
effect for the network of suppliers can be properly measured as the ratio of the root mean square errors (RMSE) of: (i) the total orders re-
ceived by the vendors in NV , fu  V tg1t0 , where

X
 V t :
u rs t;
u 8t;
8r;s2E1 [E4

 D tg1
and (ii) the total customer demands owing out of ND , fu t0 , where

1
As a special case, when there is no information sharing, the set of local information available to supplier i 2 NS should be

I0i t : fxi t; xi t  1; . . . ; xi 1; yi t; yi t  1; . . . ; yi 1g [ fuis t  1; uis t  2; . . . ; uis 1 : i; s 2 Eg [ furi t  1; uri t  2; . . . ; uri 1 : r; i 2 Eg:
2
Improper policies will amplify inventory perturbations over time and result in unbounded costs (Ouyang and Daganzo, 2006a). Intuitively, most practical ordering policies
chosen by rational suppliers should satisfy the properness assumption.
802 Y. Ouyang, X. Li / European Journal of Operational Research 201 (2010) 799810

X
 D t :
u rs t;
u 8t:
8r;s2E1 [E2

The bullwhip effect is measured by


P1 
Denition 1. When customer demands are known and non-steady 2
t0 uD t 0 , the supply network described by (6)(8) experiences
the bullwhip effect of magnitude
P1 2 12
 V t
u
Pt0
1 : 9
2
u
t0 D t

When the customer demands are not known, however, the worst-case RMSE amplication factor, W, across all possible customer demand
sequences will be used:

Denition 2. A supply network described by (6)(8) is said to experience the bullwhip effect if

P1 2 12
 V t
u
W : sup Pt0
1 > 1: 10

8fuD tg0 2
u
t0 D t

This worst-case metric and its variants are proposed and discussed in (Daganzo, 2001; Ouyang and Daganzo, 2006a) and also used in
(Ouyang, 2007; Ouyang and Daganzo, 2008). The condition W 6 1 guarantees that the RMSE is not amplied (i.e., that the bullwhip effect
does not arise) under any (unknown) customer ordering scenario. This metric is suitable for designing a robust supply chain that avoids the
bullwhip effect no matter what the customer does. It also has the advantage of scalability. A large supply network can essentially be par-
titioned into multiple subnetworks for analysis; with metric (10), the elimination of the bullwhip effect in each subnetwork ensures that
the entire supply network avoids the bullwhip effect.

3. Frequency domain analysis and results

Frequency domain analysis has been used to study properties of time series for a long time; see (Brockwell and Davis, 1998) for a review.
The standard techniques (e.g., orthogonal decomposition, transformation, and superposition) signicantly facilitate the quantication of a
linear systems response to input signals. Interested readers are referred to Ouyang and Daganzo (2009) for a review of recent applications
of such techniques to the bullwhip effect in serial supply chains. This section shows how a similar framework can be adopted to simplify
the analysis of complex supply chain networks.

3.1. When customer demands are known and stationary

The supply network can be regarded as a multiple-input, multiple-output (MIMO) system, with all customer demands as the inputs, and
all supplier orders as the outputs. Similar to those in (Ouyang and Daganzo, 2006a, 2008), the metrics (9) and (10) can be expressed equiv-
alently in the frequency domain. Any realization of customer demand fu  rs tg1 8r; s 2 E1 [ E2 g can be decomposed by a discrete Fou-
p;
t0
rier transform into a set of pure harmonic components, Ars we jwt
(where j 1), each with an angular frequency w 2 p; p and a
complex amplitude Ars w 2 C. Because Eqs. (6)(8) are linear and time-invariant, it is known that for any harmonic component of any
customer demand (as input of the system), Ars wejwt , the resulting orders placed by each supplier of the linear network (as output of
the system) are also harmonic with the same frequency but a different complex amplitude, Ars ik we
jwt
; 8i; k 2 E3 [ E4 . The actual order
P
sequence, fu  ik tg1
t0 , is the superposition of the harmonic components r;s2E1 [E2 A rs
ik wejwt
across frequency domain w 2 p; p. As a
result, the total order received by the vendors, fu  V tg1
t0 , is the superposition of (i) Aik we
rs jwt
; i; k 2 E4 ; r; s 2 E1 [ E2 and (ii)
Ars we jwt
; r; s 2 E1 across frequency w 2 p; p.
Ars w
We can dene an jE3 j jE4 j  jE1 j jE2 j transfer function matrix, T : T rs rs jw
ik , such that T ik e Ars w ; 8i; k 2 E3 [ E4 ; r; s 2
ik

E1 [ E2 . As in (Ouyang and Daganzo, 2006a), the transfer functions between any two order sequences can be derived by applying the z-
transform to the system dynamics Eqs. (6)(8). Denote the z-transforms of the error sequences by X i z : Zf xi tg; Y i z :
Zfyi tg; 8i 2 NS , and U rs z : Zfu  rs tg; 8r; s 2 E, and then for all i; k 2 E3 [ E4 ,

ZfArik Pxr tg Arik z1 X r z; ZfBrik Py


r tg Brik z1 Y r z; 8r 2 NS ; and ZfC rsik Pu rs t  1g C rsik z1 z1 U rs z; 8r; s 2 E:

Thus, if we now apply the z-transform to both sides of Eqs. (6)(8), 8i; k 2 E3 [ E4 , we obtain:

X   X  rs 
U ik z Arik z1 X r z Brik z1 Y r z z1 C ik z1 U rs z ; 11
r2NS r;s2E
X X
z  1X i z U is z  U ri z zxi 0; 12
s:i;s2E r:r;i2E
X X
z  1Y i z zlis U is  i 0:
U ri z zy 13
s:i;s2E r:r;i2E

We have assumed that the system starts from the equilibrium state at t 0, and thus  i 0 u
xi 0 y  rs 0 0; 8i; 8r; s. Algebraic
manipulations of these equations reveal that for i; k 2 E3 [ E4 :
Y. Ouyang, X. Li / European Journal of Operational Research 201 (2010) 799810 803

X   X  s 1  X  
zz  1U ik z zArik z1 z1lrs Brik z1 U rs z  zAik z zBsik z1 U rs z z  1C rs 1
ik z U rs z
r2NS ;r;s2E s2NS ;r;s2E r;s2E
X
rs
a
ik zU rs z; 14
r;s2E

where
8
>
> z  1C rs 1
ik z ; if r; s 2 E1 ;
>
< z  1C rs z1  zAs z1  zBs z1 ;
ik ik ik if r; s 2 E2 ;
arsik z : r 1lrs r rs 1 s s
>
> zA z 1
z B z 1
z  1C 1 1
ik z  zAik z  zBik z ; if r; s 2 E3 ;
>
: ik ik
rs 1 r 1 1lrs r 1
z  1C ik z zAik z z Bik z ; if r; s 2 E4 :
rs
Note that a ik z gives the inuence of U rs z on U ik z. For notation convenience, we index the arcs in E1 by e1 ; e2 ; . . . ; ejE1 j , those in E2 by
ejE1 j1 ; ejE1 j2 ; . . . ; ejE1 jjE2 j , those in E3 by ejE1 jjE2 j1 ; ejE1 jjE2 j2 ; . . . ; ejE1 jjE2 jjE3 j , and those in E4 by ejE1 jjE2 jjE3 j1 ; ejE1 jjE2 jjE3 j2 ; . . . ; ejEj . Accord-
ingly, we let U d z U rs z if ed r; s 2 E. Stacking U 1 z; U 2 z; . . . ; U jEj z into vectors, we dene
2 3 2 3
U 1 z U jE1 jjE2 j1 z
6 .. 7 6 .. 7
U0 z : 6
4 .
7;
5 Uz : 6
4 .
7:
5
U jE1 jjE2 j z U jEj z

The system of Eq. (14) can be summarized into a matrix form:

zz  1E Uz Mz Uz dz U0 z; 15

where E is the jE3 j jE4 j  jE3 j jE4 j identity matrix, and


2 jE1 jjE2 j1
3 2 3
ajE 1 jjE2 j1
z    ajEj
jE1 jjE2 j1 z
jE1 jjE2 j
a1jE1 jjE2 j1 z    ajE1 jjE2 j1
z
6 7 6 7
6 .. .. .. 7 6 .. .. .. 7
Mz : 6 . . . 7; dz : 6 . . . 7: 16
4 5 4 5
jE1 jjE2 j1 jEj jE1 jjE2 j
a jEj z  a jEj z a1jEj z  a jEj z

The entries in Mz and dz satisfy

amd z arsik z; if em r; s 2 E and ed i; k 2 E3 [ E4 :


When zz  1E  Mz is invertible, we can write (15) into:
Uz Tz  U0 z; 17
where Tz is an jE3 j jE4 j  jE1 j jE2 j transfer function matrix,
2 3
T 1jE1 jjE2 j1 z    T jE
jE1 jjE2 j
1 jjE2 j1
z
6 7
6 .. .. .. 7 1
Tz : 6 . . . 7 zz  1E  Mz dz; 18
4 5
T 1jEj z  jE1 jjE2 j
T jEj z

whose elements show how suppliers transform the customer demands U0 z into supplier orders Uz. By assumption, the policies of all sup-
pliers are stable in time (i.e., proper). This happens if all the poles of all elements in Tz are within the unit circle of the complex plane,
fz : jzj < 1; z 2 Cg.
When the customer demands are known, their frequency domain representation U0 can be obtained by the z-transform or the Fourier
transform. If the customer demand is ergodic, the z-transform of any order stream from supplier i to supplier k, U ik z, obtained from (17),
fully characterizes all the statistics of that order stream fu ik tg. Since variance equals mean square error for an innite ergodic sequence,
Parsevals Theorem (Ogata, 1987) yields that:
 ik tg1
Theorem 1. If the customer demands are known and ergodic, the variance of order stream fu t0 , 8i; k 2 E3 [ E4 , is
Z p
1
jU ik ejw j2 dw; 19
2p p

where U ik z is determined from (17).


Theorem 1 yields the variance formula for every supplier order sequence. In terms of the bullwhip effect for the whole network, denote
 V tg1
the z-transforms of fu  1
t0 and fuD tgt0 as
X X
U V z : U rs z; and U D z : U rs z:
8r;s2E1 [E4 8r;s2E1 [E2

Based on (18), the input demands U0 z are transferred to U V z as follows:


U V z TV z  U0 z;
where the 1  jE1 j jE2 j transfer function matrix to U V z is
804 Y. Ouyang, X. Li / European Journal of Operational Research 201 (2010) 799810

" #
X
jE4 j X
jE4 j
jE j
X
jE4 j
jE j1
X
jE4 j
jE jjE j
TV z : 1 T 1jE1 jjE2 jjE3 jd z; . . . ; 1 T jE11 jjE2 jjE3 jd z; T jE11 jjE2 jjE3 jd z; . . . ; T jE11 jjE22 jjE3 jd z : 20
d1 d1 d1 d1

Note too that

U D z : 1  U0 z;

where 1 is a 1  jE1 j jE2 j row vector of 1s. Then we have

Theorem 2. If the customer demands are known and ergodic, the magnitude of the bullwhip effect in the supply network, as dened by (9), is
given by
"R p #1=2 "R p

#1=2
jU V ejw j2 dw
TV ejw  U0 ejw
2 dw
p p
Rp 2
Rp 2
: 21
p
jU D ejw j dw p
j1  U0 ejw j dw
In the above results, the customer demands fu  rs tg1
t0 ; 8r; s 2 E1 [ E2 , have been interpreted as a realization of ergodic stochastic pro-
cesses. More generally, if they are regarded as stochastic processes themselves, then as long as they are stationary, the frequency domain
representations fU d ejw g; d 1; . . . ; jE1 j jE2 j, may be regarded as stochastic processes over the continuous real interval p; p. The terms
in the previous formulas, e.g., jTV ejw  U0 ejw j and j1  U0 ejw j in (21), shall be replaced by their expectations.
According to the WienerKhintchine Theorem (Couch, 2001), the power spectral density functions and cross-spectral density functions
for these stationary customer demand processes are given by
h

2 i X
1
E
U d ejw
Cd 0 2 Cd s cosws; 8d 1; . . . ; jE1 E2 j; 22
s1
X
1
E U d1 ejw U d2 ejw  Cd1 ;d2 sejws ; 8d1 ; d2 1; . . . ; jE1 E2 j; 23
s1

where Cd s : E u  rs tu rs t  s, for ed r; s, is the auto-covariance function of process fu  rs tg1t0 with lag s, and
d1 ;d2
C s : E u
 rs tu  rs tg1
 ik t  s, for ed1 r; s and ed2 i; k, is the cross-covariance function for processes fu  1
t0 and fuik tgt0 with lag
s. These auto- and cross-covariance


functions

 are
known properties

 of the customer demand processes and can be directly used to calculate
the expected norms, E
TV ejw  U0 ejw
and E
1  U0 ejw
, and hence the order variances and the bullwhip effect magnitude. The results
on order variances and the bullwhip effect shall continue to hold for any stationary customer demands, as stated in the following theorem:
Theorem 3. Formulas (19) and (21) apply to all stationary (even non-ergodic) customer demands, with the integrands interpreted as their
expectations (which can be calculated based on (22, 23)).

3.2. When customer demands are not known

Formula (21) is exact but requires knowledge of the set of customer demand processes, U0 . Previous studies on serial chains have re-
vealed that the bullwhip effect can still be tested based on the worst-case metric (10) even when the customer demands are not known.
This subsection will show how these ndings can be generalized to supply networks.
The LTI network can be considered a multiple-input-single-output (MISO) system with inputs fu  rs t : r; s 2 E1 [ E2 g and output
 V tg. This system can be regarded as the superposition of multiple single-input-single-output (SISO) systems, each of which has
fu
 rs t; r; s 2 E1 [ E2 as the only input. It is known that for each such SISO system, the maximum amplication of RMSE (L2 gain) is equal
u
to the maximum norm of the corresponding transfer function with z ejw (i.e., the H1 norm). Obviously, the maximum RMSE amplication
of the MISO system (across all possible combinations of fu  rs t : r; s 2 E1 [ E2 g) is the largest H1 norm of all SISO systems. Thus, we have
the following lemma.
Lemma 1. The worst-case RMSE amplication for the supply network satises

W sup TV ejw 1 ; 24
8w2p;p

where k  k1 is the L1 norm of a vector.


Formula (24), which involves optimizations over continuous real intervals, is often computationally more convenient than formula (10),
which involves a set of innite-dimensional realizations. According to (24), the bullwhip effect arises if and only if

sup TV ejw 1 > 1: 25
8w2p;p

Formula (25) can be tested numerically by plotting all elements of TV ejw as functions of w and check if one of the peaks exceeds 1.
Alternatively, we have the following closed-form sufcient condition for a supply network to incur the bullwhip effect.
Theorem 4 (Sufcient condition). The LTI supply network described by (6)(8) experiences the bullwhip effect if kg1k1 > 0, where
 2 2
dTV z dTV z d TV z
gz : 1  Diag   2
: 26
dz dz dz
Y. Ouyang, X. Li / European Journal of Operational Research 201 (2010) 799810 805

Proof. See Appendix A.

4. Numerical examples

The modeling framework and analysis results presented in the previous sections are applicable to supply networks with any general
topology. Zipkin (2000) discussed several basic supply system structures beyond serial chains. One of the most common network topolo-
gies is the tree-shaped distribution system structure, where one central warehouse serves multiple downstream distributors and even
more retailers. The distributors and retailers may be located in different geographical regions to serve demands from customers in different
markets. For illustration purposes, we will focus on a few variants of distribution systems and only two types of policies: (i) order-up-to
with the up-to-level adjusted based on a moving-average of orders received (as the example in Section 2); and (ii) generalized kanban
that partly depends on in-stock inventories. The analytical formulas and tests (e.g., those in Theorems 1 and 4) will be veried by simu-
lations with customer demands following independent or cross-correlated stationary AR(1) processes.

4.1. Spatially-correlated customer demands

We consider a simple distribution network with the shape of a binary tree, as shown in Fig. 2. Then
E1 : ; E2 : f1; 3; 2; 4g; E3 : f3; 5; 4; 5g; E4 : f5; 6g:
Suppose the lead time equals 2 everywhere.
Suppose suppliers 3 (e.g., retailer) and 5 (e.g., wholesaler) use an order-up-to policy with no sharing information, where and the up-to-
level is determined by moving-average of orders received in the two most recent periods. Supplier 4 (e.g., retailer) uses a generalized kan-
ban policy as presented in (Ouyang and Daganzo, 2006a). Mathematically, the policies can be written as the following:
 35 t x3 t u
u  13 t  2; 8t:
 13 t  1 u 27
   
u45 t x4 t=8  y4 t=8 u24 t  1=2 u 24 t  2=2; 8t: 28
 56 t x5 t u
u  35 t  1 u
 45 t  1
u  45 t  2;
 35 t  2 u 8t: 29
The frequency domain representations of these orders (as labelled on the arcs in Fig. 2) can be stacked into the following vectors:
2 3
  U 3 z
U 1 z 6 7
U0 z : ; Uz : 4 U 4 z 5:
U 2 z
U 5 z
It can be easily veried that
2 3 2 3
z 0 0 2z  z1 0
6 7 6 7
Mz 4 0  18 z z1 0 5; dz 4 0 3
4
z  12 z1 5;
2z  z1 2z  z1 z 0 0
and the transfer function matrices are
2 3
2z1  z3 0
6 6z2 4 7
Tz 6 0 8z3 7z2 1
7
4 5
6z2 42z1 z3
2z1  z3 2 8z3 7z2 1

and
 
6z2  42z1  z3
TV z 2z1  z3 2 ; 3 2
: 30
8z  7z 1

Fig. 2. A simple binary supply network.


806 Y. Ouyang, X. Li / European Journal of Operational Research 201 (2010) 799810

Note that
kg1k1 maxf16; 29g > 0:
According to Theorem 4, the bullwhip effect is predicted to exist.
When customer demands are known, the variances of all order streams as well as the exact magnitude of the bullwhip effect can be
calculated. Assume that the uctuations in the customer demands, faced by the two retailers, are represented by the following two
AR(1) processes.
u  13 t e1 t g1 e3 t;
 13 t 1 q  u 8t; 31
u24 t 1 q  u
  24 t e2 t g2 e3 t; 8t; 32
for q 2 1; 1, g1 ; g2 2 f1; 0; 1g, with white noises e1 t; e2 t; e3 t with mean 0 and variance r . A variety of auto- and cross-correlations
2

among the customer demands can be represented by the selection of parameters. For example, the two demands may be independent of each
other (e.g., g1 g2 0), positively correlated (e.g., g1 g2 1), or negatively correlated (e.g., g1 g2 1, probably due to retailer com-
petition). In general, it can be veried that the auto-covariance and cross-covariance functions are
   
1 g21 r2 qjsj 1 g22 r2 qjsj g g r2 qjsj
C1 s ; C2 s ; and C1;2 s 1 2 2 ;
1  q2 1q 2 1q
and the spectrum density functions, according to (22) and (23), are
     
E jU 1 ejw j2  1 g21 Uw; q; r; E jU 2 ejw j2  1 g22 Uw; q; r; E U 1 ejw U 2 ejw  E U 2 ejw U 1 ejw
g1 g2 Uw; q; r; 33
where
" #
r2 X1
r2
Uw; q; r : 12 qs cosws :
1q 2
s1
1  2q cosw q2

Thus, from (30) and (33), the inner products in the numerator and denominator of (21), after taking expectations, can, respectively be ex-
panded into the following:
Z p Z p
E jTV ejw  U0 ejw j2 dw 1 g21 j2ejw  e3jw j4 Uw; q; r dw
p p
Z p " 2jw #
6e  42ejw  e3jw 2ejw  e3jw 2
2g1 g2 R Uw; q; r dw
p 8e3jw  7e2jw 1
Z

  p
6e2jw  42ejw  e3jw
2
1 g22

Uw; q; r dw;

8e3jw  7e2jw 1

p

and
Z p h iZ p
E j1  U0 ejw j2 dw 2 g1 g2 2 Uw; q; r dw;
p p

where R is the real component of a complex number. From Theorem 3, the magnitude of the bullwhip effect is given by the ratio of these
integrals.

5
RMSE amplification

1=2=1,simulated
1=2=1,formula
2 1=2=0,simulated
1=2=0,formula
1=2=1,simulated
1=2=1,formula

0
0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8

Fig. 3. Simulation results and theoretical predictions for the simple supply network.
Y. Ouyang, X. Li / European Journal of Operational Research 201 (2010) 799810 807

The analytical results such as (21) can be easily veried by numerical simulations. Fig. 3 plots the RMSE amplication with
q 2 0:9; 0:9 and three combinations of gs. It is clear that the calculated RMSE amplications overlap exactly with simulated values
for all cases, and the bullwhip effect does exist as Theorem 4 predicts. The gure also depicts how the bullwhip effect is inuenced by tem-
poral (auto-) and spatial (cross-) correlations between the customer demands. When the temporal auto-correlation varnishes (i.e., q ! 0),
the RMSE amplication reaches the peak. On the other hand, the spatial cross-correlation between the customer demands also affects the
bullwhip effect. For any given value of q, the RMSE amplications decrease uniformly when the cross-correlation varies from negative
g1 g2 to positive g1 g2 . This implies that potential competitions over customer demands tend to exaggerate the bullwhip
effect.

4.2. Retailer collaborations via demand diversion

Now we consider a vendor (vendor 6) serving two markets through two independent wholesalers (suppliers 5, 50 ) and their enfranchised
retailers (suppliers 3 and 4, 30 and 40 , respectively); see Fig. 4. For now, ignore the four dashed arcs, then
E1 : ; E2 : f1; 3; 2; 4; 10 ; 30 ; 20 ; 40 g; E3 : f3; 5; 4; 5; 30 ; 50 ; 40 ; 50 g; E4 : f5; 6; 50 ; 60 g:
Again, the lead time equals 2 everywhere.
Suppose suppliers 3, 4 and 5 use the same order-up-to policy (in the form of (27) or (29)), suppliers 30 , 40 and 50 use the generalized
kanban policy (in the form of (28)). Due to symmetry, the order stability performance of this network will be similar to that of the previous
example.
Now suppose that the retailers in the same market (although belonging to different wholesalers) start to collaborate. In order to balance
the received demand, the suppliers in market k 1; 2 agree to divert a small fraction dk 0 6 dk  1 of received customer demand to the
other retailer. Assume that the diversion of customer demands happens right after receipt of orders, and such diversion is achieved by
expedite orders (i.e., with zero lead time) as illustrated by the four dashed arcs in Fig. 4. Then, the augmented network now has
E3 : f3; 5; 4; 5; 30 ; 50 ; 40 ; 50 ; 3; 30 ; 30 ; 3; 4; 40 ; 40 ; 4g:
The relevant expedite ordering policies are
u  13 t;
 330 t d1  u  30 3 t d1  u
u 1030 t;  440 t d2  u
u  24 t;  40 4 t d2  u
u 2040 t:
We note that the demand diversion is not conducted when d1 d2 0.
For this system, similar matrix manipulation shows that the transfer function matrix to U V z is
"
4d1 2z2  3z4 6z2  42 4d1 2z2  3z4
TV z 2z1  z3 2 d1 z4  2z2  ;  d1 z4  2z2 ; 2z1  z3 2
8z3  7z2 12 8z3  7z2 12
#
4 2 4d2 2z2  3z4 6z2  42 4d2 2z2  3z4 4 2
d2 z  2z  ;  d2 z  2z :
8z3  7z2 12 8z3  7z2 12

Without knowing the customer demand, the bullwhip effect is predicted to still exist, because
 
kg1k1 max 16  2d1 4d21 ; 42 2d1 4d21 ; 16  2d2 4d22 ; 42 2d2 4d22 > 0:
However, we will see that the collaboration among retailers may mitigate the bullwhip effect, as demonstrated in the following scenarios.

Fig. 4. A vendors distribution network serving two markets.


808 Y. Ouyang, X. Li / European Journal of Operational Research 201 (2010) 799810

RMSE amplification
4

3
1= 2=1, 1=2=0
1= 2=1, 1=2=0.25
2 1= 2=1, 1=2=0.5
1=2=0, 1=2=0
1=2=0, 1=2=0.25
1=2=0, 1=2=0.5
1 1=2=1, 1=2=0
1=2=1, 1=2=0.25
1=2=1, 1=2=0.5
0
0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8

Fig. 5. The effect of demand diversion.

Suppose the customer demands in each market are represented by the same cross-correlated AR(1) processes as those in Section 4.1;
i.e., demands fu  24 g follow (31) and demands fu
 13 g; fu  10 30 g; fu
 20 40 g follow (32). The magnitude of the bullwhip effect is calculated from for-
mula (21) and estimated from simulations for a range of parameters. The results are shown in Fig. 5, where each curve represents over-
lapping theoretical and simulated results. Again, we see similar relationships between the RMSE amplication and various demand
correlations. Interestingly, when the values of d1 and d2 (i.e., representing the extent of retailer collaborations) increase in the selected
range, the RMSE amplications decrease in all customer demand scenarios, and the reduction is the most signicant for the case of negative
cross-correlation. This is intuitive; the diversion operation enables retailers to stabilize their received demands by averaging-out the uc-
tuations. Negative cross-correlation in demands provides larger potential to cancel out the uctuations. This is why demand diversion re-
duces the bullwhip effect especially when competitions are prevailing.

5. Conclusions and future research

This paper has presented a system control framework for analyzing order stability and the bullwhip effect in complex supply networks.
We developed exact formulas for the variances of the order streams anywhere in the network based on system operating policies and the
spectrum of stationary customer demand processes. This also yields the exact magnitude of the bullwhip effect. We also derived robust
analytical conditions to predict whether or not the bullwhip effect will arise, without knowing the customer demands. The numerical
examples have shown how the presented framework enables us to study the effect of various factors (e.g., network structure, spatial
and temporal customer demand correlations, supplier operating strategies) on the bullwhip effect.
The results of this paper generalized previous ndings by accommodating general network topologies, multiple customers, and more
general demand processes. This work provides the basis for modeling complex interactions among suppliers (such as competitions and
collaborations) and among customer demands (such as spatial correlations across markets). The analysis framework presented in this pa-
per can be further generalized by relaxing the linear and time-invariant assumption. One direction is to allow for nonlinear (probably
piece-wise linear) ordering policies and stock-based lead times. This will prevent negative orders and negative inventory. Another direction
is to consider possible stochasticity in system characteristics such as production yields, transportation delay, and other operating mecha-
nisms (e.g., dynamic pricing and supplier competitions).

Acknowledgment

The nal part of this research was supported by the US National Science Foundation through Grant CMMI-0748067. The helpful com-
ments of the anonymous reviewers are gratefully acknowledged.

Appendix A. Proof for Theorem 4

The proof for Theorem 4 relies on the following lemma.


Lemma 2. For any proper and LTI SISO system with transfer function Tz, 9x 2 p; p such that jTejx j > 1 if T1 1 and

2

d Tz

2
dTz

dTz

 
> 0: 34
dz
z1 dz
z1 dz

2
z1

Proof. When Tejw is analytical for all w, Taylor expansion in the neighborhood of w 0 gives
Y. Ouyang, X. Li / European Journal of Operational Research 201 (2010) 799810 809




(

)
dT ejx

d Tz

2 2
jx dTejx

x2 3 dTz

dTz

x2
Te T1 x
Ox 1 j x
Ox3 ; 35
dx
w0 dx2
2 dz
z1 dz
z1 dz

2 2
w0 z1

Note that the second term of Eq. (35) is purely imaginary. Hence,
"(

# )2 
2
d Tz

2
dTz

x2 dTz

jTejx j2 1
O x 3
x O x 3

dz
z1 dz
z1 2
2 dz
z1
(
2

)
d Tz

2
dTz

dTz

1

 2

x2 Ox3
dz z1 dz z1 dz
z1
h
i2

dTz
dTz
d2 Tz

Hence, if dz

 dz

dz2 z1
> 0, by continuity of jTejx j, there exists x  0 such that jTejx j > 1. This completes the proof. h
z1 z1
It should be noted from the proof of Lemma 2 that the property of jTejx j is investigated only in the neighborhood of x 0. Hence, it
provides a sufcient but not necessary condition for the bullwhip effect to exist. In other words, there are cases where Eq. (34) is violated
but jTejx j > 1 at some x  0.
More specically, Lemma 2 can be expressed in the following form.
Pn
bi zi P P
Corollary 1. If the transfer function is expressed in the canonical form, Tz Pi0
n , such that a0 1 and ni0 ai ni0 bi , then there exists
a zi
i0 i
jx
x 2 p; p with jTe j > 1 if
1 Xn X n X n
Pn 3 j  iai bj k  i  jak kbk  > 0:
i0 ai i0 j0 k0

In light of Lemma 2, Theorem 4 can be proven by observing that each element in TV z is the SISO transfer function from the correspond-
ing demand to the output, U V z. Therefore, according to Lemmas 1 and 2,

W sup kTV ejw k1 > 1


8w2p;p

if at least one transfer function element in TV z, i.e., Tz, satises (34). Then, simple algebraic manipulation directly leads to Theorem 4.

References

Baganha, M.P., Cohen, M.A., 1998. The stabilizing effect of inventory in supply chains. Operations Research 46 (3), 572583.
Blinder, A.S., 1986. Can the production smoothing model of inventory behavior be saved? Quarterly Journal of Economics 101, 431454.
Boccadoro, M., Martinelli, F., Valigi, P., 2006. H-innity control of a Supply Chain model. In: Proceedings of the 45th IEEE Conference on Decision and Control, pp. 43874392.
Brockwell, P.J., Davis, R.A., 1998. Time Series: Theory and Methods, second ed. Springer, New York.
Cachon, G., Randall, T., Schmidt, G.M., 2007. In search of the bullwhip effect. Manufacturing and Service Operations Management 9 (4), 457479.
Chen, F., Drezner, Z., Ryan, J., Simchi-Levi, D., 2000a. Quantifying the bullwhip effect in a simple supply chain: The impact of forecasting, lead times, and information.
Management Science 46 (3), 436443.
Chen, F., Ryan, J., Simchi-Levi, D., 2000b. The impact of exponential smoothing forecasts on the bullwhip effect. Naval Research Logistics 47 (4), 271286.
Cooke, J.A., 1993. The $30 Billion promise. Trafc Management 32, 5759.
Couch, L.W., 2001. Digital and Analog Communications Systems, sixth ed. Prentice-Hall, New Jersey.
Daganzo, C.F., 2001 A theory of supply chains. Institute of Transportation Studies Research Report, UCB-ITS-RR-2001-7, University of California, Berkeley, CA.
Daganzo, C.F., 2003. A Theory of Supply Chains. Springer, Heidelberg, Germany.
Daganzo, C.F., 2004. On the stability of supply chains. Operations Research 52 (6), 909921.
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.
Disney, S.M., Farasyn, I., Lambrecht, M., Towill, D.R., Van de Velde, W., 2006. Taming the bullwhip effect whilst watching customer service in a single echelon of a supply chain.
European Journal of Operational Research 173, 151172.
Forrester, J., 1958. Industrial dynamics, a major breakthrough for decision makers. Harvard Business Review 36, 3766 (July and August).
Forrester, J., 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.
Gavirneni, S., Kapuscinski, R., Tayur, S., 1999. Value of information in capacitated supply chains. Management Science 45 (1), 1624.
Gilbert, K., 2005. An ARIMA supply chain model. Management Science 51 (2), 305310.
Goodwin, J., Franklin, S., 1994. The beer distribution game: using simulation to teach systems thinking. Journal of Management Development 13 (8), 715.
Graves, S., 1999. A single item inventory model for a nonstationary demand process. Manufacturing and Service Operations Management 1, 5061.
Holt, C.C., Modigliani, F., Muth, J., Simon, H.A., 1960. Planning Production, Inventories and Work Force. Prentice-Hall, Englewood Cliffs, New York.
Kahn, J., 1987. Inventories and the volatility of production. American Economic Review 77, 667679.
Kaminsky, P., Simchi-Levi, D., 1998. The Computerized Beer Game: Teaching the Value of Integrated Supply Chain Management. In: Lee, H., Ng, S.M. (Eds.), Supply Chain and
Technology Management, The Production and Operations Management Society.
Lee, H.L., Padmanabhan, V., Whang, S., 1997a. The Bullwhip effect in supply chains. Sloan Management Review 38 (3), 93102.
Lee, H.L., Padmanabhan, V., Whang, S., 1997b. Information distortion in a supply chain: The bullwhip effect. Management Science 43 (4), 546558.
Lee, H.L., So, K.C., Tang, C.S., 2000. The value of information sharing in a two level supply chain. Management Science 46 (5), 628643.
Magee, J.F., 1956. Guides to inventory control (Part II). Harvard Business Review, 106116.
Magee, J.F., Boodman, D., 1967. Production Planning and Inventory Control, second ed. McGraw-Hill, New York.
Naish, H.F., 1994. Production smoothing in the linear quadratic inventory model. Economic Journal 104 (425), 864875.
Ogata, K., 1987. Discrete-time Control Systems. Prentice-Hall, New Jersey.
Ouyang, Y., 2005 System-level Stability and Optimality of Decentralized Supply Chains, Ph.D. Dissertation, University of California, Berkeley.
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., Daganzo, C.F., 2006a. Characterization of the bullwhip effect in linear, time-invariant supply chains: some formulae and tests. Management Science 52 (10), 1544
1556.
Ouyang, Y., Daganzo, C.F., 2006b. Counteracting the bullwhip effect with decentralized negotiations and advance demand information. Physica A 363 (1), 1423.
Ouyang, Y., Daganzo, C.F., 2009. Robust stability analysis of decentralized supply chains. In: Kempf, K., et al. (Eds.), Production Planning Handbook, Springers International
Series in Operations Research and Management Science, Springer, in press.
Ouyang, Y., Daganzo, C.F., 2008. Robust tests for the bullwhip effect in supply chains with stochastic dynamics. European Journal of Operational Research 185 (1), 340353.
810 Y. Ouyang, X. Li / European Journal of Operational Research 201 (2010) 799810

Ouyang, Y., Lago, A., Daganzo, C.F., 2006. Taming the bullwhip effect: from trafc to supply chains. In: Carranza, O., et al. (Eds.), Bullwhip Effect in Supply Chains: A Review of
Methods, Components, and Cases, Palgrave-MacMilan Series.
Ramey, V.A., 1991. Nonconvex costs and the behavior of inventories. Journal of Political Economy 99, 306334.
Sterman, J., 1989. Modelling managerial behaviour: misperceptions of feedback in a dynamic decision making experiment. Management Science 35 (3), 321339.
Zhang, X., 2004. Evolution of ARMA demand in supply chains. Manufacturing Service and Operations Management 6, 195198.
Zipkin, P.H., 2000. Foundations of Inventory Management. McGraw-Hill/Irwin.

S-ar putea să vă placă și