This article was downloaded by: [155.33.16.

124] On: 13 October 2014, At: 07:44

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

Operations Research
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

The Stabilizing Effect of Inventory in Supply Chains

Manuel P. Baganha, Morris A. Cohen,

To cite this article:

Manuel P. Baganha, Morris A. Cohen, (1998) The Stabilizing Effect of Inventory in Supply Chains. Operations Research 46(3-
supplement-3):S72-S83. http://dx.doi.org/10.1287/opre.46.3.S72

Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

© 1998 INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 THE STABILIZING EFFECT OF INVENTORY IN SUPPLY CHAINS
MANUEL P. BAGANHA
Universidade Nova de Lisboa, Lisbon, Portugal

MORRIS A. COHEN
University of Pennsylvania, Philadelphia, Pennsylvania
Downloaded from informs.org by [155.33.16.124] on 13 October 2014, at 07:44 . For personal use only, all rights reserved.

(Received November 1990; revisions received November 1995, May 1996, October 1996; accepted October 1996)

This paper presents a hierarchical model framework for the analysis of the stabilizing effect of inventories in multiechelon manufac-
turing/distribution supply chains. This framework is used to compare the variance of demand to the variance of replenishment orders
at different echelons of the system (e.g., retailer, wholesaler). Empirical data and experience with management games suggest that, in
most industries, inventory management policies can have a destabilizing effect by increasing the volatility of demand as it passes up
through the chain (the “bullwhip” effect). Our model helps to explain these observations and indicates mechanisms that can promote
stabilization. The analysis results also define sufficient conditions for the existence of stabilization and relate these conditions to the
optimality of myopic control policies.

O ne of the principal reasons used to justify investment

in inventories is its role as a buffer to absorb demand
variability. The need for such a buffer is based on strong
economic incentives for reducing fluctuations in produc-
tion levels (due to fixed setup costs and/or convex manu-
facturing costs), as well as protection against stockouts
which can affect revenue. Inventories can act to protect the
production system from changes in product requirements
generated by the marketplace. If this stabilizing function is
performed effectively, observed variability in production
will be less than the observed variability in market
demand.
There is considerable empirical evidence, however, that
indicates that the variance of retail sales is in fact less than
the variance of deliveries from plants to retailers.1 These
data suggest that inventories do not have a stabilizing ef-
fect on material flow patterns in spite of the significant
investment in finished product inventories in many indus-
tries.2 There is also evidence that the variance of plant
production is greater than the variance of manufacturer
sales (West 1986, Blanchard 1983). Furthermore, fluctua-
tions of orders tend to increase as one moves upstream in
a supply chain (Holt et al. 1968). Finally, extensive experi-
ence with management simulations (such as the Beer
Game) demonstrate how easily this “bullwhip” effect can
be triggered (Sterman 1989). Taken together, these empir-
ical results suggest that the manufacturing/distribution
supply chain can act to amplify market variability rather
than to dampen it.
Such behavior is inconsistent with management expec-
tations (which are based on the economies and technology
of manufacturing). A number of normative models of
materials management have been developed which help
to explain how various factors (like batch sizing, safety
stocks, price speculation, demand patterns) affect the
pattern of replenishment requests. In this paper, we de-
velop a material management model to analyze the stabi-
lizing effect of inventories in distribution networks. Our
model considers the important effects of multiple echelons
(wholesale, retail, market) in an environment of nonsta-
tionary (correlated) product requirements. Such nonsta-
tionarity is a consequence of the stock replenishment
procedures commonly used in each echelon of the distri-
bution network.
The model introduced in this paper can shed light on
the empirical results, noted above, concerning demand
variance amplification. It also provides a basis for develop-
ing inventory control policies for multiechelon supply
chain systems in a manner that explicitly accounts for the
linkage between stocking policies and transferred demand
processes. The model indicates that, under certain condi-
tions, inventories can have a stabilizing effect, and in other
cases this effect will not occur. These results suggest that
the stabilizing effect of inventories is based on the interac-
tion of cost, technology, and market attributes that are
peculiar to the industry of interest. They also indicate that
there is a close relationship between stabilization and the
inventory control policies utilized at each echelon of the
supply chain.
The paper is organized as follows. We begin by review-
ing the empirical data that describe the effect of invento-
ries on product requirements variance in the
manufacturing/distribution supply chain. In Section 2, rel-
evant literature is reviewed. We then present our multiec-
helon model. We use the model to analyze the relationship
between inventory levels at each echelon in the supply
chain and, in particular, the variance of input and output
demand/order processes that are transmitted from echelon
to echelon.

Subject classifications: Multiechelon inventory systems. Inventory/production smoothing.

Area of review: MANUFACTURING OPERATIONS.

Operations Research 0030-364X/98/4603-S072 $05.00

Vol. 46, Supp. No. 3, May–June 1998 S72 q 1998 INFORMS
 BAGANHA AND COHEN / S73
Table I
Observed Variability of Production and Sales
Manufacturer Wholesaler Retail
Production Sales1 Sales2 Sales3
Current Prices
Mean 161,294 161,308 91,590 88,603
Downloaded from informs.org by [155.33.16.124] on 13 October 2014, at 07:44 . For personal use only, all rights reserved.

St. Deviation 22,319 22,217 17,115 15,682

Coef. Variation 0.138 0.138 0.187 0.177
1967 Prices4
Mean 62,499 62,463 35,178 33,657
St. Deviation 4,119 3,623 2,556 1,758
Coef. Variation 0.066 0.058 0.073 0.052
1
Manufacturer shipments represent net selling values, f.o.b. at the plant, less all discounts and allow-
ances and exclude freight charges and excise taxes. The time series of these shipments reflects the
demand for goods and services which are produced by U.S. manufacturers. Manufacturers inventories
are at book value of stock on hand at the end of the period (month) and include raw materials and
supplies, work in process, and finished goods. All values are seasonally adjusted by the Bureau of Census.
2
The series on wholesale sales and inventories is limited to merchant wholesalers. It represents the
dollar sales volume of merchandise sold to retailers by wholesalers less returns, allowances, and dis-
counts, and receipts from repairs.
3
Retail sales include merchandise sold for cash or credit by establishments primarily engaged in retail
trade and thus represents the market demand for products.
4
Since these sales series are expressed in current dollars, it is necessary to deflate the data in order to
reflect actual material flows. We used the producer price and consumer price indices (also published by
the Joint Economic Committee based on data compiled by the Department of Labor) for this purpose.
Producer Price Indices measure average changes in prices received in primary markets of the U.S. by
producers of commodities at all stages of processing. Consumer Price Indices provide a measure of price
change for fixed market baskets of goods and services purchased by all urban consumers. Both indices
have as basis, the year 1967. Since there is no specific index for intermediate transactions (wholesale), we
deflated the wholesalers sales series by the Producer Index.

1. EMPIRICAL RESULTS
Previously published data on the occurrence of stabiliza-
tion does not reflect the impact of those economic agents
that operate between producers and retailers. Such agents
are found in the echelons associated with the producer’s
owned distribution network (comprising distribution cen-
ters, warehouses) or in echelons operated by independent
wholesalers and distributors who purchase and resell man-
ufactured products. In order to account for this more com-
plex (and more realistic) structure, we must explicitly
consider data for the case where the distribution network
has multiple echelons.
The “Economic Indicators,” published by the Joint Eco-
nomic Committee of the U.S. Congress (1978 –1985), con-
tain statistics on shipments from manufacturers,
wholesaler sales, and retailer sales. This publication sum-
marizes monthly estimates compiled by the Bureau of
Census. We use these three time series of sales to repre-
sent the production/distribution behavior of the economy
with respect to the three major supply chain echelons
(manufacturing, distribution, retail). The results are sum-
marized in Table I. The estimates in each of these time
series were developed from representative samples for
each industry which contained data from firms of all sizes
and in all kinds of businesses. None of these time series
represents actual production volumes. In order to estimate
production schedules, we used the manufacturer ship-
ments series and corrected it by the change in manufac-
turer inventory levels. Since the shipments and inventory
series reflect different item values (market price versus
cost), it is necessary to convert inventories from a cost to
market value basis. We used the approach introduced by
West (1983).3
As a result of the indicated calculations, time series on
(1) retail sales, (2) wholesale sales, (3) manufacturer sales,
and (4) production volumes were generated. We computed
the mean, standard deviation, and coefficient of variation
for each of the time series, at both current and 1967 prices,
for the period January 1978 to December 1985. The results
are presented in Table I.
Analysis of Table I shows that, at current prices, the
standard deviation of wholesaler sales is greater than that
of retail sales and smaller than that of manufacturer sales.
Manufacturer sales standard deviation is itself smaller than
the standard deviation of production. We note these re-
sults do not take into account the relative size of the ma-
terial flows in each time series.4 To take size into
consideration, we computed the coefficient of variation.
From Table I we see that, at current prices, the coefficient
of variation increases as we go from the retailer to the
wholesaler level (as noted by Blinder 1981a, 1981b). This
measure of variation then decreases in going from the
wholesaler to manufacturer level. Finally, we note that the
coefficient of variation of production is less than the coef-
ficient of variation of retail sales. Similar results are ob-
tained when we consider the 1967 prices version of the
series. However, in this case, the reduction on the coeffi-
cient of variation from the wholesaler to manufacturer
 S74 / BAGANHA AND COHEN
level is not sufficient to offset the increase from retail to
wholesaler level, and the coefficient of variation of produc-
tion is greater than that of manufacturer sales.
These results clearly support the conclusion of Blinder
and others that the variation in deliveries to retailers from
wholesalers (0.073) is greater than the variation of the
market demand (0.052), and West’s results, which indicate
Downloaded from informs.org by [155.33.16.124] on 13 October 2014, at 07:44 . For personal use only, all rights reserved.
that the variation of production (0.066) is greater than the
variation of manufacturers shipments (0.058). However,
they also highlight a fact not previously noted. The vari-
ability of manufacturers’ sales is less than the variability of
wholesaler sales, and it may even be less than the variabil-
ity of retail sales (under current prices in our data set).
This suggests that wholesalers perform a stabilizing role by
smoothing the flow of demands. Without the existence of
wholesalers, it is clear that the variability of manufacturer
shipments would be considerably larger— e.g., equal to the
variance of wholesaler shipments.
In the remainder of this paper, we develop a normative
model of distribution system management which can shed
light on the above results. This model specifically considers
a production/distribution system with three echelons, the
retailer, distribution center and producer. The model will
indicate that the observed pattern of material flow is com-
patible with rational behavior.
2. LITERATURE REVIEW
Economists have long recognized the relevance of inven-
tory fluctuations in explaining the timing and magnitude of
business cycles. The linear decision rule of Holt et al.
(1960) (HMMS) has been the basis of much theoretical
and empirical work and indicates that production smooth-
ing (the production time series has less variance than the
demand time series) is associated with cost minimizing
behavior. Their model indicates, in particular, the role of
buffer inventories in smoothing production in face of ran-
dom or seasonal sales. In the context of a distribution
system, the HMMS model tradeoffs would predict the
smoothing of replenishment orders from retailers by
wholesalers.
Blinder’s results, noted above, do not agree with the
conclusions suggested by the HMMS model. Blinder
(1981a) advanced the use of (s, S) policies by retailers as a
possible explanation for the observed behavior pattern. He
developed an econometric model, consistent with this class
of stock control policies.
The empirical results noted above motivated several
studies concerned with the stabilizing effect of inventories.
Abel (1985) considered a model in which a firm must de-
cide, each period, how much to produce and at what price
to sell its output (monopolist situation). He assumed that
there is a one-period production lag. He uses as a measure
of stabilization the derivative of production with respect to
initial inventory. If the derivative is between 21 and 0,
inventories will have a stabilizing effect. Abel showed that
if the demand curve is perfectly inelastic and if there are
lost sales, then stabilization occurs. He also developed a
relationship between this concept of stabilization and the
variance related definition we introduced earlier. Kahn
(1985) considered the case where, in each period, a firm
must decide how much product to make available for sale.
He showed that the presence of positive serial correlation
in the demand process and the use of backlogging for
excess demand both lead to a situation where the variance
of production is greater than the variance of sales. His
results are also valid for the monopolist price setter.
Caplin (1985) showed that, for a retailer following an
(s, S) continuous review inventory policy, the variance of
replenishment orders exceeds the variance of demand.
This result is also valid when aggregated over several re-
tailers. Thus, the variance of aggregate orders in a distri-
bution network is greater than the variance of aggregate
demand. Caplin’s result (which pertains to a stationary
steady-state model) suggests that the presence of retailers
does not act to stabilize the aggregate demand faced by the
wholesaler who supplies them.
Lee et al. (1994) developed a framework for explaining
the destabilization effect which is based on four different
mechanisms: (1) order batching, (2) price speculation, (3)
capacity shortfalls that lead to over-ordering and cancella-
tion, and (4) demand signal processing (whereby the mar-
ket demand signal is distorted through the customer re-
order process). They devised a set of simple models to
illustrate how each of these factors can lead to the ampli-
fication of order variance as one moves up a production/
distribution supply chain. The model presented here
relates specifically to the order batching and demand sig-
naling phenomena.
It must be emphasized that Abel (1985) and Kahn
(1985) consider only price and production costs (no inven-
tory related costs), while Caplin’s results are independent
of the firm’s cost structure (and hence optimal stocking
policies are not considered). Caplin’s model is stationary,
while Abel and Kahn explicitly consider the dynamic as-
pects of production/inventory requirements. In this paper,
we present a multiechelon model that combines the rele-
vant features of the models noted above. It contains an
explicit model of inventory stock control at each echelon
and location and, moreover, nonstationary demand and
supply processes are considered.
3. THE NORMATIVE MODEL
In this section we develop a multiechelon model that can
explain the empirical results presented in Section 1. This
model formulation takes into account the multiechelon
linkages. Our main results focus on an explanation of the
stabilizing role of wholesalers.
3.1. Model Description
Consider the network structure illustrated in Figure 1. We
assume a periodic review system. At the lowest level, re-
tailers face a random i.i.d. market demand process. At the

Downloaded from informs.org by [155.33.16.124] on 13 October 2014, at 07:44 . For personal use only, all rights reserved.
Figure 1. Multiechelon structure.
beginning of each period, retailers review their inventory
positions. Depending on their stock control policy and
their current status, an order may be issued, which will be
delivered after a fixed leadtime. All excess demand is
backlogged.
The retailers are all supplied by one distribution center,
which reviews its inventory position at the beginning of
each period. The distribution center will place a replenish-
ment order to the plant according to its inventory control
policy (which is to be determined).
3.2. Assumptions
The following assumptions are made in our model.
Assumption 1. Demand per period at each retailer location
is an independent and identically distributed random vari-
able.
Assumption 2. A periodic review procedure is used, and in
each period the following sequence of events occurs at each
stocking location: order, delivery and sale.
Assumption 3. Orders are delivered after a fixed leadtime.
Assumption 4. Excess demand is backlogged at all levels.
Assumption 5. Fixed plus variable ordering costs are in-
curred at the retailer level. Variable costs only are charged
at the distribution center.
BAGANHA AND COHEN / S75
Assumption 6. Holding and shortage costs are charged
against expected (end of period) inventory levels.
Assumption 7. Pipeline holding costs are paid by the re-
ceiving location.
Assumption 8. All costs are stationary.
Assumption 1, concerning stationarity of demand,
means that the only source of requirements serial correla-
tion is the retailers’ ordering decisions. Periodic review
and backlogging are common practices in many industries.
The assumed ordering cost structure charges fixed replen-
ishment costs to retailers and not to the wholesaler. This
situation could arise when wholesalers have adopted effi-
cient order filling technology and/or where the manufac-
turers absorb the cost of filling orders. In the latter case, it
could be argued that such costs are passed on to the
wholesaler and are reflected in the unit cost (and hence in
the holding cost). We have observed such arrangements
when wholesalers have long term contracts with their sup-
pliers.
3.3. Notation
We make use of three subscripts in the model. The first
subscript refers to the echelon, the second to the location,
and the third to the time period. Thus Xijt will be the value
of variable X in period t, at location j of echelon i. For
stationary parameters, we omit the time period.
The echelons will be numbered starting from the bot-
tom. (Retailers will be echelon 1.)
For each location, echelon and period, define:
D: a random variable denoting demand in a period;
Q: the quantity of stock ordered;
I: the amount of inventory on hand at the beginning
of a period;
W: the inventory position (on hand plus on order) at
the beginning of a period;
c: unit cost of ordering ($/unit);
K: fixed cost of ordering ($/order);
h: unit holding cost ($/unit/period);
p: unit shortage cost ($/unit/period);
L: leadtime (in periods);
f[: probability density function for D;
F[: cumulative distribution function for D;
m: mean demand;
s: variance of demand;
b: discount factor (0 ¶ b ¶ 1).
3.4. Formulation
The managerial objective is to determine order quantities
for each period and location which minimize total ex-
pected discounted cost over an infinite horizon.5 There are
two cost components in the system (retailers—CR and dis-
tribution center—CDC). The model statement is as fol-
lows.
HO J
N
Min E CR i 1 CDC (1)
i51
 S76 / BAGANHA AND COHEN
subject to

CR i 5 Ob
`

t51
t21
H K 1i d ~Q 1it ! 1 c 1i Q 1it

S
1 b L 1i h 1i W 1it 1 Q 1it 2 O
t1L 1i

l5t
D 1il D1
Downloaded from informs.org by [155.33.16.124] on 13 October 2014, at 07:44 . For personal use only, all rights reserved.

1 b L 1i p 1i SOt1L 1i

l5t
D 1il 2 W 1it 2 Q 1it DJ1
, (2)

CDC 5 Ob
`

t51
t21
H c 21 Q 21t

S
1 b L 21 h 21 W 21t 1Q 21t 2 OD
t1L 21

l5t
21l D
1

1 b L 21 p 21 SO
t1L 21

l5t
D 21l 2W 21t 2Q 21t DJ
1
,

(3)
W 1it11 5 W 1it 1 Q 1it 2 D 1it , (4)
W 21t11 5 W 21t 1 Q 21t 2 D 21t , (5)

OQ
N
D 21t 5 1it , (6)
i51

d ~ x! 5 $ 10 if x . 0,
if x < 0,
0 < b < 1,
Q 1it > 0, i 5 1, . . . , N,
Q 21t > 0.
Equation (2) is the total discounted cost over an infinite
horizon for retailer i. It includes the usual costs for an
inventory system with backlogging, fixed order cost, and
leadtime. We are assuming that the ordering costs are
incurred in the order period and that its variable compo-
nent includes the holding cost for the pipeline inventory.6
Equation (3) corresponds to the distribution center cost
and is similar to Equation (2) without the fixed order cost.
Equations (4)–(6) are the inventory balance equations
across time (Equations (4) and (5)) and locations (Equa-
tion (6)). The latter balance equation captures the demand
linkages between echelons. These linkages will, in turn,
determine the degree of stabilization that will occur across
echelons in the network.
4. SOLUTION METHOD
We solve the overall optimization problem by a multiech-
elon decomposition procedure. Each location will deter-
mine its optimal policy only after all those locations it
supplies have determined their inventory policies and
their associated ordering policies. The diagram in Fig-
ure 2 illustrates this procedure. At the lower level, retailers
will determine their optimal stocking policy. Once this is
done, the demand transmitted to the distribution center
can be calculated. This demand acts as an input to the
(7) Figure 2. Multiechelon decomposition.
(8)
(9)
distribution center for the determination of optimal distri-
(10) bution stock control policy. We then move up to the next
level. Note that this process is not necessarily optimal from
the point of view of the overall network—a branch-and-
bound procedure would be required (see Cohen et al.
1986). However, if each location corresponds to a different
firm, then this solution approach provides a reasonable
approximation to a decentralized control mode. We also
assume that there are no interactions between levels (e.g.,
lateral pooling).
We note that the inventory policy of the distribution
center will have an impact on the leadtime offered for
replenishment orders made by the retailers. This, in turn,
can impact the retailer inventory policy. The multiechelon
decomposition proceeds only once from the bottom up. A
more realistic, but computationally demanding, approach
would cycle through the echelons until policy convergence
occurs. Experimentation by Cohen and Lee (1988) for sim-
ilar manufacturing/distribution supply chain optimization
problems indicates fast convergence and good first itera-
tion accuracy for this procedure. Recent results by Bag-
anha and Cohen (1990) show that the cost performance of
policies generated in the first iteration of the multiechelon
decomposition procedure is comparable to that attained
with the final iteration (fixed point) policy. In the following
subsections, we will solve the optimization problem for
each echelon based on the assumption of a one-pass pro-
cedure.
 BAGANHA AND COHEN / S77

O F ~ j!,
4.1. The Retailer Subproblem `
~n!
M i ~ j! 5 i
We define the stocking problem at the retailer level in n51

m ~ j! 5 O f ~ j!,
terms of the following mathematical program: `
~n!

HO J N i i
n51
Min E CR i , (11)
i51 are, respectively, the renewal function and the renewal
Downloaded from informs.org by [155.33.16.124] on 13 October 2014, at 07:44 . For personal use only, all rights reserved.

subject to density associated with the demand process. It follows

H
from (17) that the variance of a retailer’s orders will be
Ob
`
CR i 5 t21
K 1i d ~Q 1it ! 1 c 1i Q 1it greater than the variance of its input market demand.
t51 Analysis of values of ri( j) for different values of j shows

S D
that the successive values of Q1it are not independent.7 It is
OD
t1L 1i 1
1 b L 1i h 1i W 1it 1 Q 1it 2 1il important to note that this correlation is negative for a lag
l5t
of one period. For higher order lags, it might be either

1 b L 1i p 1i SOt1L 1i

l5t
D 1il 2 W 1it 2 Q 1it DJ
1
,
positive or negative. This follows from the batch ordering
enforced by (s, S) policies.
In what follows we assume that the reorder process can
(12) be described by an autoregressive model with lag parame-
W 1it11 5 W 1it 1 Q 1it 2 D 1it , (13) ter less than or equal to the cycle length, which is defined

$ 10 %
if x . 0, as the number of periods between two consecutive orders
d ~ x! 5 (14) and is given by the expression:
if x < 0,
0 < b < 1, (15) cycle length p i 5 1/P@Q 1it . 0#.
Q 1it > 0, i 5 1, . . . , N. (16)
Thus for any location i,
Since there are no constraints linking the N retailers
O u ~ j!Q
pi
(i.e., we have assumed there is no pooling or coordination Q 1it 5 a i 1 i 1it2j 1 u it , (18)
among the retailers), the problem decomposes into the j51

solution of N stochastic dynamic inventory problems with where uit is a random disturbance term that is assumed to
backlogging and a fixed order cost. It is well known that be a normal variate with mean 0 and variance s2.
the optimal policy for each of these problems is of (s, S) To calculate values for the autoregressive model param-
type (e.g., Scarf 1960, Veinott 1966) and there are algo- eters (ui( j)), we use the Yule-Walker equations (the loca-
rithms to compute it (Veinott and Wagner 1965, Feder- tion index i has been suppressed) that relate the
gruen and Zipkin 1984, Cohen et al. 1988). autocorrelation coefficients to these parameters:
Define (si, Si) as the parameter values for the optimal
policy at location i. Let ri( j) be the correlation coefficient r~1! 5 u ~1! 1 u ~2!r~1! 1···
between the order quantity in period t, Q1it, and the order 1 u ~ p!r~ p 2 1!
quantity in period t 1 j, Q1it1j, i.e., the autocorrelation for r~2! 5 u ~1!r~1! 1 u ~2! 1···
the retailer replenishment request time series for a lag of j 1 u ~ p!r~ p 2 2!
periods. Based on a renewal approximation introduced by
: : : : : : :
Erhardt et al. (1981), it can be shown that
r~ p! 5 u ~1!r~ p 2 1! 1 u ~2!r~ p 2 2!
O
1···
`
2m i j51 m i ~ j! 1 u ~ p!.
Var~Q 1ip ! 5 s i2 1 , (17)
1 1 M i ~D i 2 1!
Since the AR( p) process is stationary, there is a solution
and to the above system of equations (Box and Jenkins 1970).

O Once the us are determined we may compute ai as

5O
D i 21
2E~Q 1ip ! lf i ~l !/Var~Q 1ip !, j 5 1,
! O
l50
D i 21 a i 5 E~Q 1ip !~1 2 u ~1! 2 u ~2! 2 · · · 2 u ~ p!!. (19)
r i ~ j! 5 2E~Q 1ip l50 f i~ j!~l !/Var~Q 1ip !,
Similarly, the variance of ut is computed from its relation
2 j21
l51 F i~ j21!~D i 2 1!r i ~l !, j . 1,
with variance of Q1ip.
where
Var~u it !
Di 5 Si 2 si, Var~Q 1ip ! 5 . (20)
1 2 u ~1! 2 · · · 2 u ~ p!
Q1ip 5 a random variable denoting the steady state
order quantity, 4.2. The Distribution Center
i [ 5 the
f (j) probability function of the j-fold The demand at the distribution center is determined by
convolution of demand per period, the sum (convolution) of the replenishment orders it re-
i [ 5 the CDF of the j-fold convolution of demand,
F (j) ceives from each of the retailers. Thus, the random vari-
and able denoting demand at the distribution center is
 S78 / BAGANHA AND COHEN

OQ where (Xut 2 1) is an operator that conditions the random

N
D 21t 5 1it , variable X on the information available at the end of period
i51
t 2 1 (or beginning of t).
where Q1it is defined in (18). The total demand during periods t . . . t 1 L will be
The sum of N autoregressive processes will be an autore-
OD O OD Opf
L L p k11
gressive moving average process. Since any stationary ARMA
t1k 5 t2j i k1j112i
process may be converted into an AR process of superior
Downloaded from informs.org by [155.33.16.124] on 13 October 2014, at 07:44 . For personal use only, all rights reserved.

k50 k50 j51 i51

order, we represent the distribution center demand by
O O gp 1 O O p e
L k11 L k11

OfD 1
p i i t1k112i
D 21t 5 g 1 j 21t2j 1 et, (21) k50 i51 k50 i51

OD O Opf
j51 p L k11
where g and fj are the autoregressive model parameters 5 t2j i k1j112i
and et is the random disturbance term in period t, which is j51 k50 i51

a normal variate with mean 0 and standard deviation se.

O Op 1Oe O p,
L k11 L L2k

Note that the demand at the distribution center will also 1g i t1k i (25)
k50 i51 k50 i51
exhibit negative serial correlation for one period lag. In order
to ensure that the demand is always positive, we assume that and

SO D
the mean demand is at least m times the standard deviation L
and that realizations of demand will not deviate from the E D t1k ut 2 1
mean more than m times the standard deviation, i.e.: k50

OD O Opf O Op,
p L k11 L k11
E~D t uD t21 , . . . , D t2p ! > m s , 5 1g (26)
t2j i k1j112i i
and j51 k50 i51 k50 i51

E~D t uD t21 , . . . , D t2p ! 2 m s

< D t < E~D t uD t21 , . . . , D t2p ! 1 m s ,
Var SO L

k50
U
D t1k t 2 1 5 s e2 D k50
O S O p D.
L L2k

i51
2
i (27)

where m is an arbitrary positive number. Note that the variance is independent of the information
A key random variable for inventory control is the de- available. It is only a function of the length of the lead
mand during lead time. The following lemma (Baganha time. In the sequel we will use s(L11) for the standard
1987) defines the demand k periods into the future, given deviation of the (L 1 1) period demand. Since ¥k50 L
Dt1k
the information available now. can be expressed as a linear combination of independent
normal variates, it is a normal variate.
Lemma 1. Let demand in period t be represented by the The stock control optimization problem at the distribu-
autoregressive model of order p: tion center may now be reformulated as (ignoring location
D t 5 g 1 f 1 D t21 1 · · · 1 f p D t2p 1 e t , and echelon subscripts):
and define

p1 5 1 and pj 5 Ofp
j21
for j > 2,
Min E FO H `

t51
b t21 cQ t 1 b Lh W t 1 Q t 2 S ODD
t1L

l5t
l
1

SO D JG
i j2i
i51 t1L 1

with 1 b Lp Dl 2 Wt 2 Qt , (28)
l5t
p i2j 5 0 if i<j and fi 5 0 if i . p.
subject to
Then, given the information available at the beginning of
OfD
p
period t:
(i) the demand in period t 1 k is Dt 5 g 1 j t2j 1 et t 5 1, 2, . . . , (29)
j51

OD Opf
p k11
W t11 5 W t 1 Q t 2 D t , (30)
D t1k 5 t2j i k1j112i
j51 i51
where et is a normal variate with 0 mean and variance se2.

O gp 1 O p e
k11 k11
Using a change of variable we get the equivalent formu-
1 i i t1k112i ; (22)
i51 i51
lation.
(ii) the expected demand in period t 1 k is
Min Ob
`
t1L21
H
E cY t 1 hE Y t 2 S OD
L
D1

OD Opf O gp ;
p k11 k11 t1l
t51 l50
E~D t1k ut 2 1! 5 1 (23)
SO D
t2j i k1j112i i
j51 i51 i51 L 1
1 pE D t1l 2 Y t
(iii) the variance of demand in period t 1 k is l50

Op; S O DJ
k11 L 1
Var~D t1k ut 2 1! 5 s e2 2
i (24) 2 b cE Y t 2 D t1l , (31)
i51 l50
 BAGANHA AND COHEN / S79

OfD ÎO Op
p
L L
Dt 5 g 1 1 et, t 5 1, 2, . . . , (32) p j11
2
> j11 , (37)
j t2j
j51
j50 j50
W t 5 Y t21 2 D t21 , (33)
where
where Yt is the total inventory position in period t after

Ofp
j21
ordering but before demand.
In the following, we will first derive a myopic policy for p1 5 1 and pj 5 i j2i for j > 2,
Downloaded from informs.org by [155.33.16.124] on 13 October 2014, at 07:44 . For personal use only, all rights reserved.

i51
problems (31)–(33) and then prove a condition for opti-
mality of the myopic policy. with
The myopic solution will result from solving the one- p i2j 5 0 if i<j and fi 5 0 if i . p.
period problem,

S D
Proof. Expressions (34)–(36) define the myopic policy.
OD
L 1
Min cY t 1 hE Y t 2 t1l
This policy will be optimal if it is feasible. This translates
l50 into Qt11 Ä 0.

1 pE SO L
D t1l 2 Y t D 1
2 b cE Y t 2 S OD
L
D
1
.
From expression (36),

Op
t1l
l50 l50 L
Q t11 5 E~D t1L11 ut! 1 e t i11 .
Compute the derivative, set it to zero and solve for Yt: l50

OD O O p f O Op,
p L k11 L k11
From the conditions of the theorem and part (iii) of
Yt 5 Qs ~L11! 1 t2j i k1j112i 1g i
j51 k50 i51 k50 i51 Lemma 1, we have

where
(34)
E~D t1L11 ut! > m s ÎO L11

l51
p l2.

Q 5 Z 21 ~ h 1pp22c b c ! , Thus,
with Z being the CDF of a standard normal variate.
Equation (34) can also be written as Q t11 > m s ÎO
t50
L
p l2 1 u t Op
l50
L
l11

O E~D
ÎO
L
Y t 5 Q s ~L11! 1 t1k ut 2 1!.
Op
(35) L11 L
k50
> ms p l2 2 m s e l11 > 0.
The quantity ordered in period t 1 1, (Qt11) will be given t50 l50

by Yt11 2 (Yt 2 Dt): The sufficient condition follows from this inequality. □

O E~D O E~D
L L
Q t11 5 t1l11 ut! 2 t1l ut 2 1! 1 D t Three particular, but important, cases of Theorem 1 are
l50 l50 presented in the following corollary (the first part is imme-
O @E~D
L
diate; see corollaries 3.1 and 3.2 in Baganha 1987 for a
5 E~D t1L11 ut! 1 t1l ut! 2 E~D t1l ut 2 1!#
l51
simple proof of the second and third parts). The first part
of the corollary pertains to the case where the lead time is
1 D t 2 E~D t ut 2 1!,
zero, and the second where demand is any autoregressive
since E(Dt1lut) 2 E(Dt1lut 2 1) 5 pl11[Dt 2 E(Dtut 2 1)], model and lead time is one. The third part refers to the
situation of general deterministic lead time with demand
Op
L
Q t11 5 E~D t1L11 ut! 1 l11 @D t 2 E~D t ut 2 1!# represented by an autoregressive model of period one.
l50

OD O pf Corollary 1. If in the optimization problem defined by

p L11
5 t112j i L1j112i (31)–(33):
j51 i51

(i) L 5 0, or
O p 1e O p
L11 L
1g i t l11 . (36) (ii) L 5 1 and f1 , 0, or
i51 i50
(iii) p 5 1 and f1 , 0, then a myopic policy is optimal.
We can now present the sufficient conditions for the myo-
When a myopic policy is used, the following theorem
pic policies defined above to be optimal.
gives the sufficient condition for distribution center inven-
Theorem 1. If, in the optimization problem defined by tory to stabilize requirements passed on to the manufac-
(31)–(33): turer.
(i) Dt Ä 0, @t,
Theorem 2. A sufficient condition for the inventory stabili-
(ii) uDt 2 E(Dt)u¶ m=Var(Dt), @t,
zation effect at the distribution center following a myopic
then a sufficient condition for optimality of myopic policy is policy (i.e., Var(Q21t) ¶ Var(D21t)), is that:
 S80 / BAGANHA AND COHEN

O O
L11 L12 Table II
p i p j < 0. (38) Myopic/Stability Conditions
i51 j5i11
Myopic Optimal Stabilization
Proof. From (36), we can calculate:
L50 Always f1 , 0

Var~Q t11 ! 5 O Var~D

p
FO t112j !
L11
p i f L111j2i G 2 L51
p51
f1 , 0
f1 , 0
f1 1 f21 1 f31 1 f2 1 f1f2 , 0
f1 , 0
Downloaded from informs.org by [155.33.16.124] on 13 October 2014, at 07:44 . For personal use only, all rights reserved.

j51 i51

1 s e2FO G S O L

l50
p l11
2
12
L11

i51
p i f L122i D The above theorem and corollaries show that the intro-
duction of a distribution center between retailers and plant

z SO D
l50
L
p l11 cov~D t , e t !. (39)
may reduce the variability of the demand faced by the
plant. If there were no distribution center, the retailers
would order directly from the plant. Then the plant would
Since the process is stationary, Var(Dt11) 5 Var(Dt1L11). face a demand equal to the demand input for the distribu-
Substituting L 1 1 for k in (22) and computing the vari- tion center in our model (which would have variance
ance, we will get greater than market demand). Introduction of the distribu-

O Var~D FO G
p L11 2 tion center thus has the potential to introduce a stabilizing
Var~D t1L11 ! 5 t112j ! p i f L111j2i effect.
j51 i51
It is interesting to note that the myopic optimality con-

1 s e2 FO G L

l50
p l11
2
. (40)
dition (37) and the stabilization condition (38) at the dis-
tribution center are actually very similar. From (37) we
have that myopic is optimal if
Since cov(Dt, et) 5 s2e , Var(Qt11) ¶ Var(Dt1L11) if

FO G S O
L
p l11
2
12
L11
p i f L122i DS O D OL
p l11 <
L
p l11
2
. (41)
Op
L

j50
2
j11 > SO D
L

j50
p j11
2
. (42)
l50 i51 l51 l50
From (41) we know that stabilization occurs if
Using the fact that pL12 5 pifL122i we will get,
L11
¥i51
after some computations, the result. □

Corollary 2. If the leadtime is zero (L 5 0) and f1 , 0

Op
L

j50
2
j11 > SO D
L

j50
p j11
2
1 2 p L12 O p.
L11

i51
i (43)

then Var(Q21t) ¶ Var(D21t). Thus,

6
Proof. If L 5 0 Myopic Optimal
and
O O
L11 L12

O p ,0
fStabilization.
pipj 5 p1p2 5 f1, L11
i51 j5i11 2 p L12 i
i51
and the result follows. □
From these observations, we can see that when a myopic
Corollary 3. If demand at the distribution center is repre- policy is optimal and is followed at the distribution center,
sented by an autoregressive model of period 1 ( p 5 1) and the stabilization effect occurs under a straightforward con-
f1 , 0 then Var(Q21t) ¶ Var(D21t). dition based on the autoregressive parameters of the dis-
tribution center’s demand process (which are computable
Proof. If p 5 1 then pk 5 fk21
1 , k Ä 1, then
through the use of standard time series analysis). In Table

O O O O
L11 L12 L11 L12 II, we summarize the results of Corollaries 1, 2, and 3, and
pipj 5 f 1i21f 1j21 we also include the case L 5 1. The stabilization condition
i51 j5i11 i51 j5i11
for L 5 1 was obtained from expression (38) and using the
Of O
L11 L122i21
5 i21
f 1i1j facts that p1 5 1, p2 5 f1 and p3 5 f21 1 f2.
1
i51 j50 The relevance of a negative serial correlation is evident
from Table II for a lag of one period (f1 , 0). Our
5 O f
L11
2i21 1 2 f 1L122i
1
1 2 f1 analysis of the retailer’s stock control problem resulted in
i51
a replenishment process with these characteristics. Conse-
5
1
1 2 f1
FOf L11

i51
2i21
1 2 Of
L11

i51
L1i11
1 G quently, the input demand process for the distribution cen-
ter (which is the aggregation of the replenishment process

5 F f1 ~1 2 f12~L11!! f1L12~12 f1L11!

1 2 f12
2
1 2 f1
1
1 2 f1 G over all retailers) will exhibit such a negative serial
correlation.
It is worthwhile to study the case of zero leadtime in
f 1 ~1 2 f 1L11!~1 2 f 1L! detail, since it provides insight into the reasons for the
5 < 0. □
~1 1 f 1 !~1 2 f 1 ! 2 occurrence of stabilization. When L 5 0, from (36) and
 BAGANHA AND COHEN / S81
the fact that demand is an autoregressive process of order Table III
p, we have: Distribution Center Order Process: Identical Retailers

OfD
p Lead
D t11 5 j t112j 1 g 1 e t11 , Time (L) 0 1 2 3 4 5
j51
Mean 250.51 250.69 250.55 250.66 250.58 250.64
5O f D
p
Q t11 1 g 1 et. Std. Dev. 60.75 69.76 42.09 53.31 35.86 45.20
j t112j
Downloaded from informs.org by [155.33.16.124] on 13 October 2014, at 07:44 . For personal use only, all rights reserved.

j51

The only difference between the two expressions is the
error term (e). Both the demand in period (t 1 1) and the
order size in period (t 1 1) are equal to an estimate of the
demand (¥j51 p
fjDt112j 1 g) plus an error. However, in
the demand case the error is independent of the estimate
of the demand, while in the order size case the error is
correlated with the estimate of demand. The error is posi-
tively correlated with one of the components of the esti-
mated demand (Dt), and hence reinforces its impact. This
simple fact explains the existence of stabilization or ampli-
fication of demand shocks. If demand in consecutive peri-
ods is positively correlated (f1 . 0), there will be an
amplification of the shocks; while if demand in consecutive
periods is negatively correlated (f1 , 0), there will be
stabilization.
Time lags amplify only the effects described above. In
the case where (L . 0), the conditions get messier, since
the order size is a function of the aggregation of demand
over the leadtime. However, we conjecture that the intu-
ition is the same; i.e., order quantity equals estimated de-
mand plus an error that may or may not be correlated with
the estimate.
5. NUMERICAL EXAMPLES
In this section we consider two simulation examples that
illustrate the variance stabilization impact of distribution
center inventory.
In our first example, the supply chain has five identical
retailers, each facing a normal demand per period with a
mean 50 and standard deviation of 15. Assume all retailers
use an (s, S) ordering policy with s 5 60 and S 5 120. We
ran the simulation for 300 periods in order to generate a
time series of aggregated orders placed by the retailers.
The mean and standard deviation of this aggregated order
series are, respectively, 250.57 and 100.99.
Using standard time series analysis software (e.g., SAS),
we can fit an autoregression model to the aggregated order
series and get
D t 5 250.67 2 0.762D t21 1 e t , e t , N~0, 65.60!,
which indicates the presence of negative serial correla-
tion.8 In particular, we see that for this case p 5 1, f1 5
20.762, g 5 250.67, and se 5 65.60. Since p 5 1 and f1
, 0, it follows from Corollaries 1 and 3 that both the
myopic optimality and variance stabilization conditions are
satisfied at the distribution center. In order to simulate the
actual variance levels of the distribution center’s ordering
process, we substituted the above parameters into (34) for
a range of procurement lead times L to obtain the myopic
order-up-to-level function, assuming a safety stock factor
of Q 5 3. We then simulated the distribution center’s
inventory order response to the above aggregated retailer
order series. The resulting distribution center’s order se-
ries has the mean and standard deviations indicated in
Table III for each value of L.
We note that, in all cases, the standard deviation of the
distribution center’s orders is less than the aggregate retail
order series standard deviation of 100.99. The mean de-
mand, of course, is unchanged since we assume backlog-
ging, and hence no demand is lost to the system. It is
interesting to note that the distribution center order pro-
cess standard deviation exceeds the aggregate retail mar-
ket standard deviation of 33.5(=5 3 15) for all values of
L. Hence, the multiechelon system (consisting of both the
retail and distribution center levels) increases variability
overall.
Our second example consists of five retailers (Table IV).
The first three each face a normally distributed demand
with a mean of 50 and standard deviation of 15, and use an
(s, S) policy with s 5 60 and S 5 160. Each of the remain-
ing two retailers has a normally distributed demand with a
mean 100 and a standard deviation of 25, and uses an
(s, S) policy of s 5 120 and S 5 250. After simulation has
been run for 300 periods, the aggregated order time series
of the five retailers has a mean of 350.61 and standard
deviation 175.19.
The fitted autoregression model of the aggregated order
series in this case is
D t 5 350.72 2 1.02D t21 2 0.30D t22 1 e t ,
e t , N~0, 103.86!.
Hence, p 5 2, f1 5 21.02, f2 5 20.30, g 5 350.72, and
se 5 103.86. Substituting the relevant parameters into
(37) and (38), respectively, we can note that both the my-
opic optimality and variance stabilization conditions again
are satisfied for the range of values of L. Upon computing
the myopic order-up-to-levels in (34), we then simulated
the distribution center’s order process and derived the re-
sult illustrated in Table IV.
Table IV
Distribution Center Order Process: Nonidentical
Retailers
Lead
Time (L) 0 1 2 3 4 5
Mean 351.38 351.41 351.41 351.40 351.41 351.40
Std. Dev. 94.54 92.65 39.08 54.12 41.39 45.23
 S82 / BAGANHA AND COHEN
Table V
Supply Chain Variability (for Distribution Center L 5 4)
Manufacturer
Demand
Case 1 (Identical Retailers):
Mean 250
Downloaded from informs.org by [155.33.16.124] on 13 October 2014, at 07:44 . For personal use only, all rights reserved.
Std. Deviation 35.86
Coef. of Variation 0.143
Case 2 (Nonidentical Retailers):
Mean 351.4
Std. Deviation 41.39
Coef. of Variation 0.118
Once again, the distribution center acts to reduce vari-
ability. In this case, we note that the stabilization effect of
the distribution center is sufficient to reduce variability
passed on to the manufacturer; the variability is less than
the retail market aggregate standard deviation of 43.8 for
the cases L 5 2 and L 5 4.
The two examples introduced here illustrate that the
variance stabilization effect of a distribution center is quite
robust. The results of these examples, from a supply chain
perspective, are summarized in Table V for the case where
L 5 4. It is interesting to compare these results to the
empirical results in Table I, which are remarkably similar.
6. CONCLUSIONS
In this paper we reviewed some empirical results concern-
ing the destabilizing effect of inventories. We saw that
wholesalers can, in fact, introduce a degree of stabilization
into the supply chain by transmitting an order process to
manufacturers with variability lower than the variability
inherent in the retailer replenishment order process. We
also noted that there are no models available for the study
of the impact of inventory policies on the variability of
demand throughout the manufacturing/distribution supply
chain that takes the stochastic linkages associated with the
multiechelon structure of such systems into account.
The multiechelon model introduced here does consider
these linkages, and it suggests that reduction of demand
variability is possible. In particular, we showed that under
certain conditions the variance of the demand faced by a
plant will be less when it is filtered through a distribution
center than when it is received directly from replenishment
orders issued by retailers who act to amplify demand vari-
ance. The conditions that ensure variance reduction are
based on the pattern of orders generated by retailers and
are not dependent on wholesaler costs. They are, not sur-
prisingly, highly dependent on the leadtime facing the
wholesaler. The intriguing similarity of the myopic opti-
mality and the variance reduction conditions leads us to
conjecture that both phenomena will appear simulta-
neously. More importantly, it illustrates the strong linkage
between stabilization and wholesaler inventory control
practices. These conjectures are borne out by the results of
the numerical examples introduced in the preceding
section.
Distribution Center Retail
Demand Demand
250 250
101.0 33.5
0.404 0.134
350 350
175.2 43.8
0.501 0.125
The analytical models developed in this paper may be
extended in several directions. At the distribution center
we considered only variable ordering costs. The inclusion
of fixed ordering costs would introduce a factor that works
against variance reduction. A possible extension would be
to see how sensitive the model is to this fixed cost; how
high must it be in order to compensate for the other fac-
tors?
The optimization problem was solved by multiechelon
decomposition assuming that each location only has local
information. This assumption is reasonable if different lo-
cations are owned by different firms. When a firm owns the
entire multiechelon structure, or if suppliers and custom-
ers share data (through an EDI link), it is possible to use
global information in optimizing the network. How will
this affect the results? In particular, can the system be
managed in such a way that inventories absorb the shocks
of demand variability, or will the occurrence of stabiliza-
tion continue to depend on the wholesaler leadtime? Re-
cently, Cachon (1995) has explored these phenomena in a
model inspired by the efficient consumer response policies
used in the grocery industry.
The main objective of the model developed in this paper
is to show how the optimal behavior of firms can lead to a
stabilizing effect of inventories and to show that variance
amplification is not always present throughout the supply
chain.
7. ENDNOTES
1. Blinder (1981a, 1981b) presents results that show that
during the period of January 1959 to December 1980,
the variance about the trend of retail sales was smaller
than the variance of deliveries to retailers.
2. The term inventory stabilization will be used to charac-
terize the variability of the replenishment process for
stocking locations in a distribution network. In general,
demand fluctuations throughout the manufacturing/dis-
tribution chain are generated by the stochastic process
associated with the market demand. Stabilization occurs
at a location as a result of material control decisions
made in response to the product requirements faced by
the location. More precisely, we say that a location ex-
hibits stabilization when the variance of its inputs re-
quirements is greater than the variance of its output

requirements, which are determined by its material
management decisions. In the context of production/
inventory systems, we observe stabilization when the
variance of demand experienced by a finished goods
stockpile is greater than the variance of replenishment
orders sent by a stockpile to its supplier location (i.e., a
distribution center or the production facility). Blinder’s
Downloaded from informs.org by [155.33.16.124] on 13 October 2014, at 07:44 . For personal use only, all rights reserved.
results concerning plant deliveries can be viewed as a
proxy for plant demand.
3. West estimated the ratio of shipments to cost of goods
sold for aggregate manufacturing at the two-digit SIC
code level and for all manufacturing. This ratio is an
approximation to the ratio of cost to market price.
4. The values of the four series differ considerably. We
assume that the variability of the samples is representa-
tive of the variability of the population.
5. In this formulation, we are assuming that all the loca-
tions behave as price takers and will sell the product at
the same price. Hence, profit and revenue consider-
ations can be ignored.
6. c1i, for example, includes the holding cost associated
with Q1it units during L1i periods, discounted to the
ordering period.
7. To be precise, we must say that the correlation is be-
tween inventory levels at the end of each period. The
correlations between order quantities stem from the in-
ventory correlation. According to our solution proce-
dure, the distribution center sees only orders, so
correlation between orders is the relevant factor.
8. p is determined by SAS as the maximum significant lag
for the autocorrelation representation.
ACKNOWLEDGMENT
This research was partially supported by Grants ECS-
8406695 and DMC-8609840 of the National Science Foun-
dation. The assistance of Yunzeng Wang in preparing this
paper is acknowledged.
REFERENCES
ABEL, A. B. 1985. Inventories, Stockouts and Production
Smoothing. Review of Economic Studies. 283–293.
BAGANHA, M. P. 1987. Contributions to the Study of Stochas-
tic Linkages in Multi-Echelon Inventory Systems. Unpub-
lished Ph.D. Dissertation. Department of Decision
Sciences, University of Pennsylvania, Philadelphia, PA.
BAGANHA, M. P. AND M. A. COHEN. 1990. Decomposition
Algorithm Performance in Multi-Echelon Inventory Sys-
tems. ORSA/TIMS Meeting, Philadelphia, PA.
BLANCHARD, O. J. 1983. The Production and Inventory Behav-
ior of the American Automobile Industry. J. Political
Economy. 91, 365– 400.
BLINDER, A. S. 1981a. Retail Inventory Investment and Busi-
ness Fluctuations. Brookings Papers on Economic Activity.
2, 443–505.
BLINDER, A. S. 1981b. Inventories and the Structure of Macro
Models. Amer. Economic Rev. 71, 11–16.
BAGANHA AND COHEN / S83
BOX, G. E. P. AND G. M. JENKINS. 1970. Time Series Analysis,
Forecasting and Control. Holden Day, San Francisco, CA.
CACHON, G. P. 1995. On the Operational Implications of Con-
tinuous Product Replenishment in the Grocery Industry.
Unpublished Ph.D. Dissertation. Department of Opera-
tions and Information Management, University of Penn-
sylvania, Philadelphia, PA.
CAPLIN, A. S. 1985. The Variability of Aggregate Demand
with (s, S) Inventory Policies. Econometrica. 53,
1395–1409.
COHEN, M. A., P. R. KLEINDORFER, AND H. L. LEE. 1986.
Optimal Stocking Policies for Low Usage Items in Multi-
Echelon Inventory Systems. Naval Res. Logist. 33, 17–38.
COHEN, M. A., P. R. KLEINDORFER, AND H. L. LEE. 1988.
Service Constrained (s, S) Inventory Systems with Prior-
ity Demand Classes and Lost Sales. Mgmt. Sci. 34,
482– 499.
COHEN, M. A. AND H. L. LEE. 1988. Strategic Analysis of
Integrated Production-Distribution Systems: Models and
Methods. Opns. Res. 36, 216 –228.
ERHARDT, R. A., C. SCHULTZ, AND H. WAGNER. 1981. (s, S)
Policies for a Wholesale Inventory System. In L. B.
Schwartz (ed.) Multi-Level Production/Inventory Control
Systems: Theory and Practice. North-Holland.
FEDERGRUEN, A. AND P. ZIPKIN. 1984. An Efficient Algorithm
to Compute Optimal (s, S) Inventory Policies. Opns. Res.
34, 1268 –1285.
HEYMAN, D. AND M. SOBEL. 1984. Stochastic Models in Opera-
tions Research, Vol. II, McGraw-Hill.
HOLT, C. C., F. MODIGLIANI, J. MUTH, AND H. SIMON. 1960.
Planning Production, Inventories and the Work Force.
Prentice-Hall.
HOLT, C. C., F. MODIGLIANI, AND J. P. SHELTON. 1968. The
Transmission of Demand Fluctuations Through Distribu-
tion and Production Systems: The TV-Set Industry. Ca-
nadian J. Economics. 14, 718 –739.
KAHN, J. 1985. Inventories and the Volatility of Production.
Mimeograph. MIT, Cambridge, MA.
LEE, H., P. PADMANABHAN, AND S. WHANG. 1994. Information
Distortion in a Supply Chain: The Bullwhip Effect. Mgmt.
Sci. 43, 546 –558.
SCARF, H. E. 1960. The Optimality of (s, S) Policies in the
Dynamic Inventory Problem. In K. J. Arrow, S. Karlin,
and P. Suppes (Eds.), Mathematical Methods in the Social
Sciences, Stanford University Press, Stanford.
STERMAN, J. D. 1989. Modeling Managerial Behavior: Impli-
cations of Feedback in a Dynamic Decision Experiment.
Mgmt. Sci. 35, 3, 321–339.
U. S. CONGRESS, JOINT ECONOMIC COMMITTEE. Economic Indi-
cators, 1978 –1985.
VEINOTT, A. F. 1966. On the Optimality of (s, S) Inventory
Policies: New Conditions and A New Proof. J. SIAM
Appl. Math. 14, 1067–1083.
VEINOTT, A. F. AND H. M. WAGNER. 1965. Computing Opti-
mal (s, S) Inventory Policies. Mgmt. Sci. 11, 525–552.
WEST, K. D. 1983. A Note on the Econometric Use of Con-
stant Dollar Inventory Series. Economic Lett. 13,
337–341.
WEST, K. D. 1986. A Variance Bound Test of the Linear
Quadratic Inventory Model. J. Political Econ. 94, 374 –
401.