Sei sulla pagina 1di 24

Chapter 3

Modeling and Measuring the Bullwhip Effect

Li Chen and Hau L. Lee

Abstract The bullwhip effect is a phenomenon commonly observed in industry. It


describes how the distortion of demand information in a supply chain amplifies de-
mand variance as it moves from consumption point up the supply chain to layers of
suppliers. The bullwhip effect has been a subject of intensive research activities. Re-
searchers have tried to address questions such as: What causes the bullwhip effect?
How would different types of demand signal processing in forecasting and replen-
ishment decisions affect the bullwhip effect? Can we explain the magnitude of the
bullwhip effect in terms of the characteristics of the product and the supply chain?
What is the magnitude of the bullwhip effect in practice, how does it differ across
industries and products, and how prevalent is the phenomenon? In this chapter, we
review both theoretical and empirical research done to address these questions, as
well as research done to identify important approaches and specifications that are
necessary to correctly measure and evaluate the true extent of the bullwhip effect.

3.1 Introduction

Demand variability and uncertainty is a driver of supply chain inventory. Manag-


ing supply chains can be a challenge when demand variability and uncertainty is
high. For a company in a supply chain consisting of multiple stages, each of which
is run by a separate organization (or company), the variability of demand faced by
this company can be much higher than the variability faced by downstream stages
(where downstream stages refers to the stages closer to the final consumption of

Li Chen
Samuel Curtis Johnson Graduate School of Management, Cornell University, Ithaca, NY 14853,
e-mail: li.chen@cornell.edu
Hau L. Lee
Graduate School of Business, Stanford University, Stanford, CA 94305,
e-mail: haulee@stanford.edu

1
2 Chen and Lee

the product). The bullwhip effect refers to the phenomenon where demand vari-
ability amplifies as one moves upstream in a supply chain (Lee et al, 1997a, or
LPW). LPW described this as a form of demand information distortion. Lee et al
(1997b) further discussed the managerial and practical aspects of the bullwhip ef-
fect, giving more industry examples. The bullwhip effect phenomenon is closely re-
lated to studies in systems dynamics (Forrester, 1961; Sterman, 1989; Senge, 1990).
Sterman (1989) observed a systematic pattern of demand variation amplification
in the Beer Game, and attributed it to behavioral causes (i.e., misperceptions of
feedback). Macroeconomists have also studied the phenomenon (Holt et al, 1968;
Blinder, 1981; Blanchard, 1983).
According to LPW, the bullwhip effect has been observed extensively in many
industries. However, they provided only anecdotal references. LPW developed sim-
ple mathematical models to explain how the bullwhip effect could arise, and iden-
tified four causes: demand signal processing, order batching, price variations, and
the rationing game. Demand signal processing refers to a company using forecast
updates, and such updates would automatically lead to larger order fluctuations than
demand. Order batching refers to companies not ordering in every single time unit,
and, as a result, order variance would naturally be larger than demand variance.
Price variations result in companies making more than normal order quantities
when prices are lower than normal, and vice versa, leading to higher order fluc-
tuations. Rational game refers to companies anticipating supply shortages, and to
ensure adequate supply, exaggerate their needs through placing larger order quanti-
ties than otherwise. Hence, these models showed that the bullwhip effect could be
a result of rational decision making under limited information and myopic objec-
tives. Croson and Donohue (2006), using clever experimental setups, demonstrated
that there could also be additional behavioral causes, namely, the under-reaction to
lags and coordination stock. Under-reaction to lags refers to players ignoring inven-
tory in the pipeline when they made ordering decisions. Coordination stock refers
to players increasing their orders because they wanted to have higher safety stocks,
resulting from experiences of past delays in shipments from suppliers.
The main contention of LPW is that one needs to understand the causes of the
bullwhip effect in order to devise counter-strategies. Hence, each of the causes re-
quires a set of strategies for companies to use. These counter-strategies tend to fall
into two broad types. First, one needs to recognize the existence of the bullwhip
effect so that one is not fooled by the distorted demand information. Companies
would therefore be smarter in making capacity and inventory decisions accordingly.
Second, companies need to work, often collaboratively, to reduce the magnitude of
the bullwhip effect. One of the most commonly cited counter-strategies is informa-
tion sharing, a topic covered in other chapters of this book. In the ideal situation,
by information sharing, a company might not be misled by the distorted demand
information, and, as a result, might not bullwhip its upstream supplier.
Since the work of LPW, two streams of research have emerged: modeling
and empirical. In the modeling stream, researchers have expanded the LPW work
through more complex modeling of the demand process to show how the bullwhip
3 Modeling and Measuring the Bullwhip Effect 3

effect could arise. In the empirical stream, instead of anecdotal evidence, researchers
have tried to measure the extent of the bullwhip effect in real industry cases.
These two streams of research have reinforced each other in our deepened un-
derstanding of the bullwhip effect. Modeling research generates insights and forms
the bases of hypotheses in empirical research. Empirical research serves to confirm
or refute some of the results derived in modeling research, but it can also suggest
potential additional causes of the bullwhip effect or additional phenomena that can
lead to new modeling research. Hence, the two streams together have provided a
healthy path for both streams of research. In this chapter, we review both streams of
work.

Fig. 3.1 Illustration of infor-


mation and material flows at a
supply chain stage

In studying the bullwhip effect, we note that there have been two primary def-
initions of bullwhip effect measurement used in the literature. It is worthwhile for
us to clarify these two definitions as they affect how one interprets the results in the
literature. LPW originally described the bullwhip effect as a form of demand infor-
mation distortion. The amplification of demand variance is based on the measure of
demand variance faced by each stage in the supply chain. Hence, consider one stage
of the chain facing its own demand variance. This stage in turn makes its ordering
decision (where order can also be interpreted as production release in a manufac-
turing setting). The orders then become the demand faced by the upstream stage.
The existence of the bullwhip effect implies that the order variance is larger than
the original demand variance. This definition captures the distortion of information
flow that goes upstream (see Fig. 3.1). A second definition, used in many empirical
studies, compares the variance of order receipts (or shipments) with the variance of
sales. Sales represent the material outflow from the current stage under consider-
ation, while order receipts (or shipments) represent the material outflow from the
upstream stage, which become the material inflow to the current stage. In some
cases, the order receipt information, if not available, is inferred from the sales and
inventory data (see Blinder, 1981; Cachon et al, 2007). This definition essentially
captures the distortion of material flow that goes downstream (see Fig. 3.1).
When the upstream stage can always supply perfectly the orders placed by the
current stage, and the current stage can always satisfy the demands that it faces, then
orders and order receipts are the same, and sales and demands also coincide. In that
case, the measures of information-flow and material-flow bullwhip effects would be
4 Chen and Lee

identical. But once there are shortages in either the upstream or the current stages,
the two measures would diverge.
The bullwhip effect measurements based on these two definitions also differ in
concept. The information-flow based definition has a direct linkage to supply chain
cost because the orders issued by a stage become a driver to the upstream inven-
tory/capacity decision. Hence, the information-flow bullwhip effect is a cost driver.
In contrast, order receipts (or shipment) information is the outcome of the upstream
order-fulfillment decision process, rather than an input to the decision process.
Hence, the material-flow bullwhip effect is the consequence of the information-flow
bullwhip effect. Moreover, in the information-flow based definition, the bullwhip
effect is a result of one decision maker, i.e., the stage in question. This decision
maker observes demand, and then makes order decisions based on various struc-
tural and economic factors. In the material-flow based definition, however, there
are three decision effects involved. First, the sales data is determined by the ac-
tual demand and the on-hand inventory, where the latter is a result of the inven-
tory decisions made in previous periods. Second, as in the information-flow based
case, the unit makes order decisions, based on structural and economic conditions.
Third, the actual order receipts from the supplier are the result of the suppliers
previous production/stocking decisions, where the order receipts may not exactly
equal the orders (e.g., production shortfall, transportation constraints, etc.). In view
of these differences, we believe the information-flow based definition is more suit-
able for theoretical analysis purposes. However, we recognize the need for using
the material-flow based definition as an empirical surrogate in some cases, and thus
include a discussion of the implications of such an approximation in Sect. 3.4.
The rest of the chapter is organized as follows. In Sect. 3.2, we review the empir-
ical findings of the bullwhip effect. In Sect. 3.3, we review the literature of bullwhip
effect modeling, with an emphasis on the demand process modeling. Section 3.4
discusses various issues related to the empirical measurement of the bullwhip ef-
fect. We conclude the chapter by discussing some future research opportunities in
Sect. 3.5.

3.2 Survey of Empirical Findings

Empirical research concerning the bullwhip effect is a large literature and, rather
than giving a comprehensive review, we highlight some significant findings. There
are two classes of empirical research. The first one is closer to the field-based ap-
proach, in which a single supply chain is the unit of analysis. The demand infor-
mation in this single supply chain, often with a single class of products in focus,
is analyzed to explore the existence of the bullwhip effect and measure its magni-
tude. The second one uses secondary data of many companies, often aggregated, to
pursue statistical analysis of the bullwhip effect.
3 Modeling and Measuring the Bullwhip Effect 5

LPW started with their anecdotal observation of excessive volatility in weekly or-
ders in both Procter & Gambles diaper supply chain and Hewlett-Packards printer
supply chain.
We highlight a few sample studies that are based on single supply chains. In a
landmark teaching case, Hammond (1994) documented how Barilla SpA, the largest
pasta producer in the world, observed strong bullwhip effects. The supply chain
membersBarilla and its customersall processed demand signals, orders were
batched, and promotions were common. At one of the distribution centers (DC) of
its largest retail customer, the weekly orders placed by this DC to Barilla had a
mean of 300 quintals and a standard deviation of 227. But the weekly sell-through
at the DC (which can be viewed as shipments to the stores) had a mean of 300
and standard deviation of 60. Suppose that we define the bullwhip ratio to be
CVout /CVin , where CVout is the coefficient of variation (CV) of the outgoing orders,
and CVin is the CV of the incoming orders. The bullwhip effect exists if the bullwhip
ratio is greater than one. The bullwhip ratio at the Barilla case was 3.75, i.e., 73%
of the variation at the DC could be explained as the distortion within the supply
chain, while the remaining 27% was the variation faced by the DC. Through VMI
(vendor-managed inventory), Barilla was able to reduce the inventory at this DC by
47%, while shortage rates dropped by 7% to almost zero.
Lai (2005) also studied a single supply chain, that of a Spanish grocery chain
Sebastian de la Fuente. The study was based on monthly product-level data at the
DC, showing prices, markups, sales delivered to stores, supplies from suppliers,
inventories and promotion. The data set contained records of 3,745 products over 29
months that pass through the DC. Regression analysis by Lai (2005) demonstrates
that the bullwhip effect existed and was mainly driven by batching by the store, as
well as two behavioral causes identified by Croson and Donohue (2006).
Fransoo and Wouters (2000) studied two supply chains of convenience foods
(salads and ready-made pasteurized meals) involving four companies in The Nether-
lands. The supply chain consists of three stages: the producer, the regional DC and
retail franchisees. Using the filtered daily sales data (from March 23 until June 5,
1998), they found that the bullwhip effect was prominent across the supply chain.
The bullwhip ratios found in their study are shown in Table 3.1.

Table 3.1 Summary of bullwhip ratios of different supply chain stages


Supply Chain Stage Bullwhip Ratio (Meals) Bullwhip Ratio (Salads)

Production 1.75 1.23


Distribution Center 1.26 2.73
Retail Franchisee 1.67 2.09

Note that both Lai (2005) and Fransoo and Wouters (2000) used sales data, and so
the bullwhip ratios that they measured were based on the material flow. Hammonds
case was based on orders placed by the DC (information flow) and the sales to the
6 Chen and Lee

DC (material flow). It showed the challenges faced by empirical researchers as it is


not easy to truly measure the information distortion aspect of the bullwhip effect,
i.e., the information-flow bullwhip effect. The study of Fransoo and Wouters (2000)
also highlights how the bullwhip effect can vary across products. Hence, one has to
be careful when conducting empirical research with data aggregated over products.
Moving from the single supply chain approach, there also have been empiri-
cal accounts of the bullwhip effect that are based on two levels of aggregation of
secondary data. These studies usually used monthly or quarterly data aggregated
across various products or firms. The first level of aggregation was on time units.
In the sample studies of single supply chains described earlier, the time unit of the
data was based on the timing of order decisions. For example, it was a week in the
case of Barilla, since orders were generated on a weekly basis; and it was monthly
in the case of Sebastian de la Fuente, since the supply chain members order on a
monthly basis. The time unit used in the convenience food supply chain study was
a day, since the supply chain ordered on a daily basis, which was necessary given
the perishable nature of meals and salads. If companies orders on a monthly basis
or a quarterly basis, then the use of monthly or quarterly data may pose no poten-
tial problem. But this is rarely the case in real life. The second level of aggregation
was on products and often across firms as well, which can also be problematic on
the validity of bullwhip effect measurement. These two aggregation problems are
discussed in detail in Sect. 3.4.
Industry-based studies are anchored on data that are aggregated over products and
firms. High production volatility was found in the TV set industry (Holt et al, 1968),
retail industry (Blinder, 1981; Mosser, 1991), automobiles (Blanchard, 1983; Kahn,
1992), cement industry (Ghali, 1987), high tech consumables (Hanssens, 1998), pa-
per products (Carlsson and Fuller, 1999), semiconductors (De Kok et al, 2005),
semiconductor equipment (Terwiesch et al, 2005), and many other industries (Miron
and Zeldes, 1988; Fair, 1989). In these studies, researchers searched for explana-
tions to reconcile the bullwhip effect with the classic production-smoothing theory
that posits that the motive for keeping inventory is to smooth production variability
rather than to amplify it. One of the leading explanations is that production smooth-
ing was missing because seasonality had been excluded from the data (e.g., Ghali,
1987).
There are also studies that are at the multi-industry or economy level (Blinder,
1986; Bivin, 1996; Cachon et al, 2007). Here, there was even more extensive aggre-
gation of products and firms, as data from products and firms in different industries
have also been combined. Fine (1999) demonstrated the bullwhip effect using this
multi-industry approach, as shown in Fig. 3.2. The fluctuation of automobile pro-
duction was clearly greater than that of the GDP (mimicking the downstream stage
of the automobile industry), while the fluctuation of machine tools (mimicking the
upstream stage of the automobile industry) was even greater.
Several recent empirical studies are worth highlighting. Cachon et al (2007) used
monthly sales and inventory data from the U.S. Census Bureau of 19922004 to ex-
amine the bullwhip effect in the manufacturing, wholesale and retail sectors. Hence,
there was aggregation of time units, and aggregation across products. The bullwhip
3 Modeling and Measuring the Bullwhip Effect 7

Fig. 3.2 Illustration of industry volatility

effect analysis was based on the material flow concept. They found that if season-
ality is included in the measurement, production smoothing indeed exists in the
retail industry and in some manufacturing industries, but not in the wholesale in-
dustry. With seasonally unadjusted data: 62% of manufacturers have a bullwhip ra-
tio less than one, 86% of retailers had a bullwhip ratio less than one, while 84% of
wholesalers have a bullwhip ratio larger than one. Hence, there is empirical evidence
that while there was a tendency for companies to bullwhip the upstream suppliers,
sometimes the desire to smooth production may be even stronger, dampening the
bullwhip effect.
Bray and Mendelson (2012) reported that about two-thirds of firms bullwhip in
a sample of 4,689 public U.S. companies over 19742008. Building on the model
of a generalized order-up-to policy proposed by Chen and Lee (2009), the authors
decomposed the bullwhip effect by information transmission lead time. They found
that demand signals with shorter time notice have greater impact on the bullwhip
effect.
Bray and Mendelson (2015) further investigated the bullwhip effect and produc-
tion smoothing in an automotive manufacturing sample comprising 162 car models
and found that 75% of the sample smooth production by at least 5%, despite the
fact that 99% of the sample exhibit the bullwhip effect. They measured production
smoothing with a structural econometric production scheduling model based on the
generalized order-up-to policy. According to their structural estimation, there exist
both a strong bullwhip effect (on average, production is 220% as variable as sales)
and robust smoothing (on average, production would be 22% more variable absent
volatility penalty costs).
Shan et al (2014) investigated the bullwhip effect using data from over 1200 pub-
lic companies in China during 20022009. They found that more than two-thirds of
the companies exhibit the bullwhip effect. Their regression analysis suggests that
8 Chen and Lee

the bullwhip effect magnitude is positively associated with average on-hand inven-
tory and persistence of demand shocks, and is negatively associated with degree of
demand seasonality.
Duan et al (2015) collected a daily product-level dataset consisting of 487 indi-
vidual products from a supermarket chain in China. They found that the magnitude
of the bullwhip effect at the product level is much more significant than those mea-
sured at the firm or industry level, suggesting that product and time aggregation may
mask the bullwhip effect measurement.
Osadchiy et al (2015) investigated the systematic risk in demand for different
industries and firms in the U.S. economy, including retail, wholesale, and manufac-
turing sectors. They used sales as a proxy for demand, and defined the systematic
risk in sales as the correlation coefficient of sales change with the contemporane-
ous market return. They found that demand signal processing does not amplify the
systematic risk, however, aggregation of orders from multiple customers and aggre-
gation of orders over time can result in the amplification of systematic risk upstream
along the supply chain.

3.3 Modeling the Bullwhip Effect

There are multiple ways of modeling the bullwhip effect along a supply chain. For
example, the bullwhip effect can be modeled as a result of judgmental errors by
human decision makers (e.g., Sterman, 1989; Chen, 1999; Steckel et al, 2004; Cro-
son and Donohue, 2006). It can also be modeled as a result of suboptimal inventory
policies. Chen et al (2000a,b) showed that, when certain demand forecast methods,
such as moving average and exponential smoothing, are used to determine a (subop-
timal) inventory policy for an AR(1) demand process, the resulting order variability
exceeds demand variability. De Kok (2012) considered a two-echelon supply chain
in which the downstream demand is stationary but the retailers forecast demand us-
ing the exponential smoothing method. He quantified the bullwhip effect as a func-
tion of the exponential smoothing parameter. Dejonckheere et al (2003) modeled a
supply chain as an engineering system and studied the bullwhip effect under certain
(suboptimal) system replenishment rules. We refer the reader to Geary et al (2006)
for a survey of the control engineering approach for modeling the bullwhip effect.
While the above approaches can all account for the bullwhip effect, there is a
certain degree of arbitrariness in the assumed irrational human behaviors and sub-
optimal inventory policies. To eliminate such arbitrariness, a normative approach is
needed, whereby the decision maker is assumed to make rational decisions in opti-
mizing the system cost performance. In the remainder of this section, we shall take
this normative approach for modeling the bullwhip effect.
3 Modeling and Measuring the Bullwhip Effect 9

3.3.1 Overview and Model Setup

An optimal order quantity from a rational decision maker is a response to the supply
and demand uncertainties of a given situation. On the demand side, LPW showed
that positively-correlated demand coupled with long lead time would amplify the
order variability, while negatively-correlated demand would dampen it. On the sup-
ply side, LPW also showed that potential supply shortages would cause downstream
stages to inflate orders and thus trigger the bullwhip effect. However, random supply
disruptions may also dampen the bullwhip effect (Rong et al, 2009) and a capacity
constraint can smooth the order quantity (Chen and Lee, 2012).
The underlying cost structure also drives the order variability. For example, fixed
ordering costs, such as full truckload and machine setup costs, will lead to large
batch orders and cause the bullwhip effect (LPW; Cachon, 1999; Chen and Lee,
2012). External cost shocks, such as promotional discounts, will induce forward-
buying behavior, which again causes the bullwhip effect (Blinder, 1986; LPW).
Conversely, explicit penalty costs for order variability will force the decision maker
to smooth order quantities (Sobel, 1969; Aviv, 2007; Cantor and Katok, 2012; Bray
and Mendelson, 2015).
To mitigate the bullwhip effect, one can thus either encourage information shar-
ing among supply chain partners to reduce the supply and demand uncertainties,
or modify the supply chain cost structure to provide economic incentives for or-
der smoothing. In what follows, we highlight some bullwhip effect models that are
closely related to the study of demand information sharing.
Consider a supply chain stage for a single product. Inventory is reviewed period-
ically at this stage. Time is divided into periods of length one and indexed forwards
(i.e., t = 0, 1, 2, ...). Let Dt denote the customer demand in period t and let (> 0)
denote the mean of demand in a period. Customer demand is fulfilled immediately
if the stage has enough on-hand inventory; unmet demand is fully backlogged. Unit
holding cost h and stockout penalty cost p are assessed and charged to the stage
at the end of each period. Inventory is replenished from an upstream stage with a
constant lead time L. The upstream stage is assumed to have ample supply. For ease
of exposition, we assume that the manager minimizes the long-run average cost;
we note that assuming a discounted cost objective function would yield the same
insights (e.g., LPW).
It is known in the literature that the optimal policy for such an inventory system
is a state-dependent base-stock policy; a static base-stock policy can be viewed as
a special case of the state-dependent policy. Let St denote the state-dependent base-
stock level in period t. Under the base-stock policy, the order quantity in period t,
denoted by Ot , can be written as

Ot = St (St1 Dt1 ) = St St1 + Dt1 . (3.1)

In the above expression, we have implicitly assumed that the order quantity in
each period can be negative, such that the base-stock level is achievable in each
period. This is equivalent to assuming that excess inventory can be freely returned to
10 Chen and Lee

the supplier. We shall make this assumption throughout this chapter for tractability
reasons. We note that the chance of a negative order quantity becomes negligible
when the order mean is sufficiently greater than the order variance. Justifications for
this assumption can be found in LPW, Aviv (2003, 2007), and Chen and Lee (2009).
A few quick observations can be made based on expression (3.1). First, when
the demand process Dt is independent and identically distributed (i.i.d.), the op-
timal policy is a static base-stock policy, i.e., St = St1 for any t. It follows that
Ot = Dt1 , and hence var(Ot ) = var(Dt1 ). Therefore, there is no bullwhip effect
in such a system. Second, when the demand process Dt is not i.i.d., the optimal
base-stock policy is state dependent. In this case, var(Ot ) may be greater or less
than var(Dt1 ), depending on the covariance between St St1 and Dt1 . Below
we consider several different demand processes, to quantify the variance ratio be-
tween order and demand.

3.3.2 AR(1) Demand Process

LPW considered an autoregressive AR(1) demand process for modeling the bull-
whip effect. Specifically, the demand in a period is defined as

Dt = (Dt1 ) + t ,

where || < 1 and t is an i.i.d. normal random variable with N(0, 02 ). Let d =
(1 ). We can rewrite the above equation as follows:

Dt = d + Dt1 + t . (3.2)

To ensure the chance of a negative demand is negligible, we assume 0  d.


Under the free-return assumption discussed above, the optimal base-stock level
can always be achieved in each period. As a result, we can determine the optimal
base-stock level St by solving the following long-run average cost minimization
problem:
  L + L + 
min E h St Dt+i + p Dt+i St ,
St i=0 i=0

where (x)+= max(x, 0). The above problem is a standard newsvendor problem, and
we know the optimal base-stock level is given by
 
1 p
St = G , (3.3)
h+ p

where G() is the cumulative distribution function of the lead time demand Li=0 Dt+i .
Thus, it remains to determine to the distribution of the lead time demand.
By leveraging the recursive expression (3.2), with some algebra, we can show
that, for k 0,
3 Modeling and Measuring the Bullwhip Effect 11

Dt+k = d + Dt+k1 + t+k


= d + (d + Dt+k2 + t+k1 ) + t+k
=
k
1 k+1
=d + k+1 Dt1 + ki t+i .
1 i=0

Because t+i is an i.i.d. normal random variable, the lead time demand Li=0 Dt+i ,
conditional on Dt1 , also follows a normal distribution, with the conditional mean
and variance given as follows:
L+1
1 k (1 L+1 )
E Li=0 Dt+i Dt1 = d

+ Dt1 ,
k=1 1 1
L+1 k ki 2 2
 
 L
var i=0 Dt+i Dt1 =
0 .
k=1 i=1

Substituting the above result into expression (3.3), we arrive at


v
L+1 k L+1 )
uL+1 k 2
1 (1 u
ki
St = d + Dt1 + z0
t ,
k=1 1 1 k=1 i=1

where z = 1 (p/(h + p)), with 1 () being the inverse standard normal cumu-
lative distribution function. Under the above optimal base-stock policy, the order
quantity in period t is given by

(1 L+1 )
Ot = St St1

+ Dt1 = (Dt1 Dt2 ) + Dt1 .
1

Using the recursive expression (3.2) again, with some algebra, we can show that

var(Ot ) 2(1 L+1 )(1 L+2 )


= 1+ . (3.4)
var(Dt1 ) (1 )

The above expression provides a characterization of the bullwhip effect under the
AR(1) demand process. We note that the AR(1) demand process reduces to an
i.i.d. process when = 0. In this case, the above expression reduces to var(Ot ) =
var(Dt1 ), which is the same as what we obtained in the i.i.d. demand case. Thus,
expression (3.4) can be viewed as a generalization of the bullwhip effect result from
the i.i.d. demand case. From the expression, we observe that, if the demand pro-
cess has a positive temporal correlation, i.e., > 0, we have var(Ot ) > var(Dt1 ).
Order variability is amplified and the bullwhip effect exists in this case. On the
other hand, if the demand process is negatively correlated, i.e., < 0, we have
var(Ot ) < var(Dt1 ). Order variability is dampened instead. Moreover, when > 0,
12 Chen and Lee

it can be shown that the ratio in (3.4) is increasing in the replenishment lead time L,
suggesting that longer lead time induces greater bullwhip effect.

3.3.3 IMA(0,1,1) Demand Process

Graves (1999) considered another simple demand process for modeling the bullwhip
effect. Specifically, demand Dt is assumed to follow an integrated moving-average
IMA(0, 1, 1) process, which is defined as follows:

Dt = Dt1 (1 )t1 + t , (3.5)

where || < 1 and t is an i.i.d. normal random variable with N(0, 02 ). We note that
the demand reduces to an i.i.d. process when = 0, and that the demand process
becomes a random walk when = 1.
Let Ft1 = Dt1 (1 )t1 . From (3.5), we have Dt = Ft1 + t . Thus, Ft1
is the best linear predictor for Dt at the end of period t 1. With some algebra, we
can show that Ft1 satisfies the following equation:

Ft1 = Dt1 + (1 )Ft2 ,

which has the same form as an exponential smoothing forecast.


By leveraging the recursive expression (3.5), with some algebra, we can show
that, for k 0,

Dt+k = Dt+k1 (1 )t+k1 + t+k


= Dt+k2 (1 )t+k2 + t+k1 + t+k
=
k1
= Ft1 + t+i + t+k .
i=0

Because t+i is an i.i.d. normal random variable, the lead time demand Li=0 Dt+i ,
conditional on Ft1 , also follows a normal distribution, with the conditional mean
and variance given as follows:

E Li=0 Dt+i Ft1 = (L + 1)Ft1 ,


L+1
var Li=0 Dt+i Ft1 = (1 + k)2 02 .

k=1

Following an argument analogous to that of the AR(1) demand process, we obtain


the optimal base-stock level under the IMA(0,1,1) demand process as
3 Modeling and Measuring the Bullwhip Effect 13
v
uL+1
u

St = (L + 1)Ft1 + z0 t (1 + k)2 ,
k=1

where z = 1 (p/(h + p)). Under the above optimal base-stock policy, the order
quantity in period t is given by

Ot = St St1

+ Dt1 = (L + 1)(Ft1 Ft2 ) + Dt1 .

We note that the above optimal order quantity is the same as the adaptive ordering
policy proposed by Graves (1999). Here we have shown that this ordering policy is
the outcome of an optimal state-dependent base-stock policy.
By using the recursive expression (3.5) again, with some algebra, we can show
that
var(Ot |Ft2 )
= (1 + + L)2 . (3.6)
var(Dt1 |Ft2 )
From the above expression, we observe that when > 0, var(Ot |Ft2 ) >
var(Dt1 |Ft2 ) and the bullwhip effect exists. Moreover, when > 0, the ratio in
(3.6) is increasing in L, suggesting again that longer lead time induces greater bull-
whip effectthe same insight as shown earlier under the AR(1) demand process. It
is worth commenting that in the above expression (3.6), the conditional variances of
order and demand are used. This is because the unconditional variances of order and
demand under the IMA(0, 1, 1) demand process are both unbounded. Therefore, we
have to resort to the conditional variance measure, which captures the order and de-
mand uncertainties conditional on the most up-to-date demand forecast information.

3.3.4 General MMFE Demand Process

Chen and Lee (2009) proposed a demand model that generalizes the demand pro-
cesses discussed above. Specifically, they assume that the demand process evolves
according to the martingale model of forecast evolution (MMFE) process (Haus-
man, 1969; Graves et al, 1986; Heath and Jackson, 1994). Under the MMFE model,
the demand Dt is defined as

Dt = + ti,t , (3.7)
i=0

where ti,t is the demand information obtained in period t i with respect to de-
mand Dt . For all i 0, ti,t is an independent normal random variable with with
N(0, i2 ). Let 2 = 2 2
i=0 i . For ease of exposition, we shall assume < be-
2
low; in the case of = , the bullwhip effect results can be modified with the
conditional variance as in Sect. 3.3.3.
14 Chen and Lee

From (3.7), the best linear predictor for demand Dt at the end of period t i
(i 0) can be defined as

Fti,t = + t j,t , (3.8)
j=i

with Ftt being the actual demand Dt itself. With some algebra, we can show Fti,t =
Fti1,t + ti,t . Hence, ti,t can be viewed as the forecast revision with regard to
demand Dt made at the end of period t i, and tt (also written as t below) is the
final uncertainty resolved during period t after demand realization. An illustration
of the MMFE demand process is given in Table 3.2.

Table 3.2 Illustration of the MMFE process


Forecast revision for future demands in each period
Demand 0 1 2 3 4 5 6
D0 = + + 0
D1 = + + 0,1 + 1
D2 = + + 0,2 + 1,2 + 2
D3 = + + 0,3 + 1,3 + 2,3 + 3
D4 = + + 0,4 + 1,4 + 2,4 + 3,4 + 4
D5 = + + 0,5 + 1,5 + 2,5 + 3,5 + 4,5 + 5
D6 = + + 0,6 + 1,6 + 2,6 + 3,6 + 4,6 + 5,6 + 6
.. .. .. .. .. .. .. .. .. ..
. . . . . . . . . .

The demand information obtained at the end of period t with regard to all future
demands can be summarized in a forecast revision vector t = [t , t,t+1 , t,t+2 , ...]T ,
where t is assumed to be i.i.d. with a multivariate normal distribution N(0, ),
where the variance-covariance matrix is given by = E{ t tT }.
The above MMFE demand model generalizes many commonly used demand
models. For example, if t = [t , 0, 0, ...]T for all t, then we have an i.i.d. demand
process. If t = [t , t , 2 t , ...]T (with 0 || < 1) for all t, then we obtain the
AR(1) demand process described in Sect. 3.3.2. If t = [t , t , t , ...]T (with
0 < 1) for all t, then we obtain the IMA(0,1,1) demand process described in
Sect. 3.3.3. It can also be shown that the MMFE model is general enough to cover
the ARIMA(p, d, q) model (Box et al, 1994; Gilbert, 2005; Gaur et al, 2005), the
linear state-space model (Aviv, 2003), and the advance demand information model
(Gallego and Ozer, 2001). We refer the reader to Chen and Lee (2009) for details.
Under the MMFE demand model, we have, for k 0,
k
Dt+k = Ft1,t+k + t+ki,t+k .
i=0
3 Modeling and Measuring the Bullwhip Effect 15

Because t+ki,t+k is an i.i.d. normal random variable, the lead time demand Li=0 Dt+i ,
conditional on { : t 1}, also follows a normal distribution, with the condi-
tional mean and variance given as follows:
L L
 

E D
t+i
, t 1 = Ft1,t+i ,
i=0 i=0
L  L+1

T
var Dt+i , t 1 = ei1 ei1 ,

i=0 i=1

where ek is the unitary vector with the k-th element being one and ei1 = ik=1 ek .
Therefore, we obtain the optimal base-stock level under the MMFE demand process
as s
L L+1
T
St = Ft1,t+i + z ei1 ei1 ,
i=0 i=1

where z = 1 (p/(h + p)). Under the above optimal base-stock policy, the order
quantity in period t is given by
L
Ot = St St1

+ Dt1 = (Ft1,t+i Ft2,t+i ) + Dt1 .
i=0

From the expressions (3.7) and (3.8), with some algebra, we can show that

var(Ot ) (eL+2 )T eL+2 L+2 T


i=1 (ei ) ei
= 1+ 1 1
. (3.9)
var(Dt1 ) 2

The above expression provides a general, unifying formula for the bullwhip effect
for demand processes that can be represented by the MMFE model. For example,
the bullwhip effect result (3.4) under the AR(1) demand process can be recovered
from (3.9) by setting t = [t , t , 2 t , ]T .
When the demand process is either AR(1) or IMA(0,1,1), we have shown that
longer lead time induces greater bullwhip effect. However, from (3.9), we cannot
claim such a result without additional assumptions on the variance-covariance ma-
trix of the forecast evolution process. The general expression (3.9) indicates that
it is actually the overall forecast correlation during the lead time period that drives
the magnitude of the bullwhip effect.

3.4 Measuring the Bullwhip Effect

As discussed in the introduction, measurement of the bullwhip effect is another


important topic. In this section, we use analytical models to demonstrate various
issues that one can encounter in measuring the bullwhip effect with empirical data.
16 Chen and Lee

When one compares the variances of the order and demand, the first question
is whether a ratio or an absolute difference metric should be used. If the objective
is to determine whether the bullwhip effect exists or not, then both the ratio and
difference metrics can be used. However, if one wants to compare the bullwhip effect
for different products, the ratio metric, being unit-independent, appears to be a better
choice. For example, consider two products: one with demand variance of 10 and
order variance of 20, and the other with demand variance of 40 and order variance
of 80. With the ratio metric, the amplification ratio is 2 for both products. However,
with the difference metric, the second product has greater amplification than the first
one (40 versus 10). Furthermore, if one tries to calculate the aggregated bullwhip
effect measure over these two products (assuming the demands are independent),
the ratio would remain 2, but the difference would increase to 50 (10 + 40). The
ratio metric is thus more suitable for comparison purposes.
In some empirical studies, the standard deviation ratio metric and/or the coef-
ficient of variation ratio metric have been used (see the Barilla example given in
Sect. 3.2). We note that the standard deviation ratio and the variance ratio contain
essentially the same information for comparison purposes, as the former metric is
just a squared root value of the latter. The coefficient of variation ratio is equivalent
to the standard deviation ratio when the order mean is the same as the demand mean.
When the order mean is different from the demand mean, one needs to normalize the
variability measure based on the different mean values. For example, suppose that
only demand data from a partial set of customers are available in a distribution net-
work. Then the order mean may be greater than the demand mean. In this case, the
coefficient of variation ratio is more appropriate for measuring the bullwhip effect.
In what follows, we shall consider a single-stage model with the order mean equal
to the demand mean. Thus, the variance ratio metric is sufficient for our analysis
purposes.

3.4.1 Seasonality

Cachon et al (2007) found that including seasonality in the measurement of the bull-
whip effect leads to a much lower bullwhip effect measurement result. The following
model is used by Chen and Lee (2012) to demonstrate such an effect.
Let s0 , ..., sT 1 denote the additive seasonality with a regular cycle of T periods,
where s0 corresponds to the seasonal factor of demand D0 . Without loss of general-
1
ity, we assume that the seasonal factors are normalized, such that Ti=0 si = 0. We
can define the variability of seasonality as
T 1
1
Vs =
T s2i .
i=0

Let Dt (Dt0 ) and Ot (Ot0 ) denote the seasonal (deseasonalized) demand and order
quantity in period t, respectively. Thus, the demand in a period can be expressed as
3 Modeling and Measuring the Bullwhip Effect 17

Dt = Dt0 + s(t mod T ) .

Under the free-return assumption, the optimal base-stock level in a period is given
by
L
St = St0 + s(t+i mod T ) ,
i=0

St0
where is the state-dependent base-stock level for the deseasonalized demand pro-
cess. Therefore, the seasonal order quantity in a period is given by

Ot = St (St1

Dt1 ) = St0 St1
0 0
+ Dt1 + s(t+L mod T ) = Ot0 + s(t+L mod T ) ,

where the last equality follows from relationship Ot0 = St0 St1
0 0 . Hence, we
+ Dt1
can show that
var(Ot ) Vs + var(Ot0 ) var(Ot0 ) var(Dt1
0 )
= 0 = 1+ 0 ,
var(Dt1 ) Vs + var(Dt1 ) Vs + var(Dt1 )

where Dt0 and Ot0 are the deseasonalized demand and order quantities. Thus, if the
bullwhip effect exists in the the deseasonalized demand process, i.e., var(Ot0 )
0 ) > 0, and if the variability of seasonality dominates the deseasonalized
var(Dt1
0 ), then including seasonality in the bullwhip
demand variability, i.e., Vs  var(Dt1
effect measurement will drive the ratio close to one. Chen and Lee (2012) further
showed that the ratio may go below one when there is a capacity limit in the system.

3.4.2 Time Aggregation

Measuring the bullwhip effect based on aggregate data also prompts the question of
potential aggregation biases in the measurement. The following model is used by
Chen and Lee (2012) to investigate the effect of data aggregation on the bullwhip
effect measurement.
Consider first the time aggregation effect. Define the M-period aggregation of
M = M1 D
demand and order as Dt1 =0 t1+ and OtM = M1
=0 Ot+ , respectively. By
the relationship (3.1), it follows that

OtM = St+M1 St1 + Dt1


M
.

For most common demand models, it can be shown that limM var(Dt1 M ) = .

For example, it can be easily verified that this condition holds for an AR(1) demand
M will
process. Thus, under this condition, the variance of the aggregated demand Dt1
eventually dominate the finite variance of St+M1 St1 as M increases. Therefore,
it is straightforward to show that limM var(OtM )/ var(Dt1 M ) = 1. That is, time

aggregation has a masking effect.


Moreover, suppose that the demand follows an ARMA(1,1) process given by
18 Chen and Lee

Dt = (Dt1 ) + t + t1 ,

where > 0, + > 0. Chen and Lee (2012) showed that the bullwhip effect under
time aggregation is given by

var(OtM ) 2( + )(1 L+1 )(1 L+2 + L+1 )


M
= 1+ .
var(Dt1 ) (M/(1 M ))(1 2 )(1 + )2 2( + )(1 + )

It is easy to verify that the above ratio decreases to one monotonically as M in-
creases.

3.4.3 Product and Location Aggregation

Besides time aggregation, empirical data are also subject to product and location
aggregation. Since location aggregation is mathematically equivalent to product ag-
gregation, below we will present the analysis on product aggregation from Chen
and Lee (2012). Define the N-product aggregation of demand and order quantities
N = N D
as Dt1 n=1 t1,n and OtN = Nn=1 Ot,n , respectively, where Dt1,n is the de-
mand for product n and Ot,n is the order quantity for product n. Also, define the
aggregated base-stock level as StN = Nn=1 St,n .
Consider first the case in which the demands of the N products are spatially-
0
independent but share a common additive seasonality pattern s0 , ..., sT 1 . Let Dt,n
0
and Ot,n denote the deseasonalized demand and order quantity for product n, respec-
tively. Thus, the demand for product n in a period can be expressed as
0
Dt,n = Dt,n + n s(t mod T ) ,

where the multiplicative factor n captures the heterogenous magnitude of season-


ality across the N products. Without loss of generality, assume n 1 for all n. Also
assume the replenishment lead time is a constant of L periods for all products. The
optimal order quantity for product n is given by
0
Ot,n = Ot,n + n s(t+L mod T ) .

We can show that the aggregate bullwhip ratio is given by


N
0 0
var(OtN )
{var(Ot,n ) var(Dt1,n )}
N )
= 1 + n=1
N 2 N
.
var(Dt1 0
V
n s + var(Dt,n )
n=1 n=1

The above ratio approaches to one as N goes to infinity. Therefore, if the products
under aggregation share a common seasonal profile (for example, the Christmas
seasonality in the retail industry), including seasonality in the measurement will
3 Modeling and Measuring the Bullwhip Effect 19

drive the aggregated bullwhip ratio to one and thus mask the bullwhip effect of
individual products.
Consider another case in which the demands of the N products are spatially-
independent but all belong to the AR(1) process family. Specifically, let us assume
that the demand for product n is given by:

Dt,n n = n (Dt1,n n ) + t,n ,

where 0 n < 1 and t,n is an i.i.d. normal random variable with N(0, n ). Thus,
we have var(Dt,n ) = n2 /(1 n2 ). By setting = n in the result of (3.4), with some
algebra, we have the following:
N
1 nL+1 1 nL+2 2
var(OtN )
2n 1 n

1 n2
n
N )
= 1 + n=1 N
.
var(Dt1 2
1 nn2
n=1

Thus, if limN Nn=1 n n2 / Nn=1 [n2 /(1 n )] = 0, the above aggregate bullwhip
ratio approaches to one as N goes to infinity. This condition can be satisfied, for
example, when n = 1/(n + 1) and n = 1 for all n, which means there is an in-
creasing portion of the products with an autocorrelation coefficient close to zero.
Another example is when n = n/(n + 1) and n = 1 for all n. In this case, there is
an increasing portion of products with autocorrelation coefficient approaching one.
In both cases, the bullwhip ratios for individual products are all strictly greater than
one (because n > 0), but product aggregation can mask the severity of the bullwhip
effect of individual products.

3.4.4 Material Flow Data

So far, we have considered bullwhip effect measurement based on the variances of


order and demand (i.e., the information flow). As discussed in Sect. 3.2, in many
empirical studies, material flow data (such as shipments and sales) are used as prox-
ies for the order and demand data to measure the bullwhip effect. Chen et al (2014)
provided an analytical comparison of these two bullwhip effect measurements.
Consider a simple i.i.d. demand process. The optimal policy in this system is a
static base-stock policy S. From (3.1), it follows that Ot = Dt1 , and there is no
bullwhip effect based on the information flow data. Now let us examine the material
flow in such a system. Let M1 (t) denote the shipment from an upstream stage in
period t. Since the upstream stage is assumed to have ample supply, we have

M1 (t) = Ot = Dt1 . (3.10)


20 Chen and Lee

Let M0 (t) denote the sales to the downstream customer in period t. It can be shown
that  L  + L  +
M0 (t) = Dt + DtL1+i S DtL+i S , (3.11)
i=0 i=0

where the last two terms represent the demand backlogs in periods t 1 and t, re-
spectively. From the above expressions, the sales M0 (t) can be written as a summa-
tion of Dt and two additional random variable terms. Intuitively, one would expect
the variance of M0 (t) to be greater than the variance of Dt . However, it can be shown
that the opposite is true. Specifically, from (3.11), the following expression can be
obtained:
 L +  L + 
var(Dt )var(M0 (t)) = 2E S DtL1+i D
tL+i S 0. (3.12)
i=0 i=0

From (3.10), we know that var(M1 (t)) = var(Dt ). Thus, it follows that var(M1 (t))
var(M0 (t)). In other words, measurement based on the material flow data would
overstate the underlying information-flow bullwhip effect in such a system. Kahn
(1987) derived a similar insight in a model with zero lead time. The above result
generalizes the insight to the case with general, positive lead time. It is clear from
the expression that the shipment variance equals the sales variance only when S = 0
or S = . Thus, measurement based on the material flow data may provide a good
approximation to the underlying information-flow bullwhip effect when the base-
stock level is either (sufficiently) high or low.
When demand Dt follows an i.i.d. normal distribution N(, 2 ), it can be shown
that the expression (3.12) is unimodal in S, reaching a peak value at S = (L + 1)
(where the system service level is 50%). It can be further shown that the expression
(3.12) is decreasing in lead time L, suggesting that measurement based on the mate-
rial flow data may provide a good approximation to the information-flow bullwhip
effect when the replenishment lead time increases.
Chen et al (2014) further considered the AR(1) demand process for comparing
the bullwhip effect measurements. They showed that the autocorrelation parameter
in the AR(1) demand model has a direct impact on the measurement biases. When
the demand is moderately correlated, measurement based on the material flow data
may overstate the information-flow bullwhip effect, which generalizes the insight
from the i.i.d. demand case. However, when the demand has a strong temporal cor-
relation, the result reverses, i.e., measurement based on the material flow data would
understate the information-flow bullwhip effect. To correct the measurement biases
from the material flow data, Chen et al (2014) proposed simple debiasing methods
based on the sample autocovariances of the sales and shipment data.
3 Modeling and Measuring the Bullwhip Effect 21

3.5 Future Research Opportunities

The vast literature of modeling and empirical research on the bullwhip effect pro-
vides several important observations. First, we need to be careful to distinguish be-
tween information-flow versus material-flow bullwhip effects. Second, the magni-
tudes of the bullwhip effects of different products do differ. Third, data aggregation
over time, product, firms and industry sectors may play a role in masking the bull-
whip effect. For practical managerial insight, a single firm has to deal with bullwhip
effects for each of its products (since production and inventory are based on a prod-
uct), and the demand orders arrive or are issued in accordance with the time units
that the partners in the supply chain use as their decision points. Hence, when prod-
ucts and firms data are aggregated, the resulting aggregate measurement may not be
that meaningful, unless one is dealing with capacities that are shared by multiple
products, or in the extreme cases, resources that are shared by multiple firms.
The second and third observations above imply that much research is still needed
for a full understanding of the bullwhip effect. Richer demand models, as well as
associated analysis of the impacts of different levels of aggregation on the measure-
ment of the bullwhip effect, are needed. We have discussed some of the recent de-
velopments in this direction in Sect. 3.3 and 3.4. Most of the models assume a linear
supply chain. It would be interesting to expand the models to include more complex
distribution networks. For example, in a system where multiple customers are sup-
plied by a common supplier, how would the inventory allocation rule at the supplier
affect the bullwhip effect and its measurement? How does the inventory pooling
effect at the supplier interact with the demand variability amplification in such a
system? How does the demand correlation among the downstream customers affect
the bullwhip effect? When each customer has a different ordering process resulting
from a different demand process, decision time unit, and ordering economics, how
does this affect the bullwhip effect?
The empirical research indicated that the tendency toward demand amplification
through demand signal processing is sometimes counteracted by the tendency to-
ward production smoothing, so that the bullwhip effect may be dampened. Chen and
Lee (2009) was one of the articles that modeled both such drivers, using a simple
way of smoothing by postponing some orders to a later time. To gain a deeper under-
standing of the interactive effects of demand amplification and production smooth-
ing, more complex production smoothing models can be constructed (e.g., Bray and
Mendelson, 2015).
Finally, we need empirical research that is based on the right units (time, prod-
ucts, or firm) to truly measure the bullwhip effect. We also need empirical research
to explore the drivers or characteristics of the products and supply chains that give
rise to different bullwhip effect magnitudes. In addition, Ozer et al (2014) have iden-
tified cultural differences between China and the US on trusts and information shar-
ing. Since the bullwhip effect is closely tied to trust factors between the buyer and
the supplier, it would also be of interest to conduct behavioral research on cultural
differences in the bullwhip effect magnitudes.
22 Chen and Lee

References

Aviv Y (2003) A time-series framework for supply chain inventory management. Oper Research
51(2):210227
Aviv Y (2007) On the benefits of collaborative forecasting partnerships between retailers and man-
ufacturers. Management Sci 53(5):777794
Bivin D (1996) Bunching in the production process. Econ Letters 50:259263
Blanchard OJ (1983) The production and inventory behavior of the American automobile industry.
J Political Econ 91(3):365400
Blinder AS (1981) Retail inventory behavior and business fluctuations. Brookings Papers on Econ
Activity 2:443520
Blinder AS (1986) Can the production smoothing model of inventory behavior be saved? Quart J
Econom 101(3):431454
Box GEP, Jenkins GM, Reinsel GC (1994) Time series analysis forecasting and control, 3rd edn.
Prentice-Hall, Englewood Cliffs, NJ
Bray RL, Mendelson H (2012) Information transmission and the bullwhip effect: An empirical
investigation. Management Sci 58(5):860875
Bray RL, Mendelson H (2015) Production smoothing and the bullwhip effect. Manufacturing Ser-
vice Oper Management 17(2):208220
Cachon G (1999) Managing supply chain demand variability with scheduled ordering policies.
Management Sci 45(6):843856
Cachon G, Randall T, Schmidt G (2007) In search of the bullwhip effect. Manufacturing Service
Oper Management 9(4):457479
Cantor D, Katok E (2012) Production smoothing in a serial supply chain: A laboratory investiga-
tion. Transportation Res. Part E 48(4):781794
Carlsson C, Fuller R (1999) Soft computing and the bullwhip effect. Econ Complexity 2(3):126
Chen F (1999) Decentralized supply chains subject to information delays. Management Sci
45(8):10761090
Chen L, Lee HL (2009) Information sharing and order variability control under a generalized
demand model. Management Sci 55(5):781797
Chen L, Lee HL (2012) Bullwhip effect measurement and its implications. Oper Research
60(4):771784
Chen L, Luo W, Shang KH (2014) Measuring the bullwhip effect with material flow data: Biases
and remedies. Working Paper. Duke University
Chen YF, Drezner Z, Ryan JK, Simchi-Levi D (2000a) Quantifying the bullwhip effect in a sim-
ple supply chain: The impact of forecasting, lead times and information. Management Sci
46(3):436443
Chen YF, Ryan JK, Simchi-Levi D (2000b) The impact of exponential smoothing forecasts on the
bullwhip effect. Naval Research Logistics 47(4):269286
Croson R, Donohue K (2006) Behaviorial causes of the bullwhip effect and the observed value of
inventory information. Management Sci 52(3):323336
Dejonckheere J, Disney SM, Lambrecht MR, Towill DR (2003) Measuring and avoiding the bull-
whip effect: A control theoretic approach. European J Oper Research 147(3):567590
De Kok AG (2012) The return of the bullwhip. Review Business Econ Lit 57(3):381393
De Kok AG, Janssen F, van Doremalen J, van Wachem E, Clerkx M, Peeters W (2005) Philips
Electronics synchronizes its supply chain to end the bullwhip effect. Interfaces 35(1):3748
Duan Y, Yao Y, Huo J (2015) Bullwhip effect under substitute products. J Oper Management
36:7589
Fair RC (1989) The production-smoothing model is alive and well. J Monet Econ 24:353370
Fine CH (1999) Clockspeed: Winning industry control in the age of temporary advantage. Perseus
Books, Cambridge, MA
Forrester JW (1961) Industrial dynamics. MIT Press, Cambridge, MA
3 Modeling and Measuring the Bullwhip Effect 23

Fransoo JC, Wouters MJF (2000) Measuring the bullwhip effect in the supply chain. Supply Chain
Management 5(2):7889
Gallego G, Ozer O (2001) Integrating replenishment decisions with advance demand information.
Management Sci 47(10):13441360
Gaur V, Giloni A, Seshadri S (2005) Information sharing in a supply chain under ARMA demand.
Management Sci 51(6):961969
Geary S, Disney SM, Towill DR (2006) On bullwhip in supply chainshistorical review, present
practice and expected future impact. Int J Production Econ 101(1):218
Ghali MA (1987) Seasonality, aggregation and the testing of the production smoothing hypothesis.
Amer Econ Review 77(3):464469
Gilbert K (2005) An ARIMA supply chain model. Management Sci 51(2):305310
Graves SC, Meal HC, Dasu S, Qiu Y (1986) Two-stage production planning in a dynamic envi-
ronment. In: Axsater S, Schneeweiss C, Silver E (eds) Multi-stage production planning and
control. Lect Notes Econ Math Systems, 266. Springer-Verlag, Berlin, Germany, p 943
Graves SC (1999) A single-item inventory model for a nonstationary demand process. Manufac-
turing Service Oper Management 1(1):5061
Hammond J (1994) Barilla SpA (A). Harvard Business School Case No. 9-694-046
Hanssens DM (1998) Order forecasts, retail sales, and the marketing mix for consumer durables.
J Forecasting 17(3):327346
Hausman WH (1969) Sequential decision problems: a model to exploit existing forecasters. Man-
agement Sci 16:B93B111
Heath DC, Jackson PL (1994) Modelling the evolution of demand forecasts with application to
safety stock analysis in production/distribution systems. IIE Trans 26(3):1730
Holt CC, Modigliani F, Shelton JP (1968) The transmission of demand fluctuations through a
distribution and production system, the TV-set industry. Canadian J Econ 1(4):718739
Kahn JA (1987) Inventories and the volatility of production. Amer Econ Review 77(4):667679
Kahn JA (1992) Why is production more volatile than sales? Theory and evidence on the stockout-
avoidance motive for inventory-holding. Quart J Econ 107(2):481510
Lai RK (2005) Bullwhip in a Spanish shop. Harvard NOM Working Paper No. 06-06
Lee HL, Padmanabhan V, Whang S (1997a) Information distortion in a supply chain: The bullwhip
effect. Management Sci 43(4):546558
Lee HL, Padmanabhan V, Whang S (1997b) The bullwhip effect in supply chains. Sloan Manage-
ment Review 38(4):93102
Miron JA, Zeldes SP (1988) Seasonality, cost shocks, and the production smoothing model of
inventories. Econometrica 56(4):877908
Mosser PC (1991) Trade inventories and (S, s). Quart J Econ 106(4):12671286
Osadchiy N, Gaur V, Seshadri S (2015) Systematic risk in supply chain networks. Management
Sci. Forthcoming
Ozer O, Zheng K, Ren Y (2014) Trust, trustworthiness, and information sharing in supply chains
bridging China and the United States. Management Sci 60(10):24352460
Rong Y, Shen Z-JM, Snyder LV (2009) The impact of ordering behavior on order-quantity variabil-
ity: A study of forward and reverse bullwhip effects. Flex Services Manufacturing J 20(1):95
124
Senge P (1990) The fifth discipline: The art and practice of the learning organization. Doubleday
Currency, New York.
Shan J, Yang S, Yang S, Zhang J (2014) An empirical study of the bullwhip effect in China. Prod
Oper Management 23(4):537551
Sobel M (1969) Production smoothing with stochastic demand I: Finite horizon case. Management
Sci 16(3):195207
Steckel J, Gupta S, Banerji A (2004) Supply chain decision making: Will shorter cycle times and
shared point-of-sale information necessarily help? Management Sci 52(4):458464
Sterman JD (1989) Modeling managerial behavior: Misperceptions of feedback in a dynamic de-
cision making experiment. Management Sci 35:321339
24 Chen and Lee

Terwiesch C, Ren Z, Ho T, Cohen MA (2005) An empirical analysis of forecast sharing in the


semiconductor equipment supply chain. Management Sci 51:208220

Potrebbero piacerti anche