Supply Chain Inventory Management and the Value of Shared Information

Author(s): Gérard P. Cachon and Marshall Fisher

Source: Management Science, Vol. 46, No. 8 (Aug., 2000), pp. 1032-1048
Published by: INFORMS
Stable URL: https://www.jstor.org/stable/2661582
Accessed: 22-08-2019 15:37 UTC

REFERENCES
Linked references are available on JSTOR for this article:
https://www.jstor.org/stable/2661582?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management
Science

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 Supply Chain Inventory Management and
the Value of Shared Information

Gerard P. Cachon * Marshall Fisher

The Wharton School, The University of Pennsylvania, Philadelphia, Pennsylvania 19104
fisher@opim.wharton. upenn.edu

In traditional supply chain inventory management, orders are the only information firms
exchange, but information technology now allows firms to share demand and inventory
data quickly and inexpensively. We study the value of sharing these data in a model with one
supplier, N identical retailers, and stationary stochastic consumer demand. There are
inventory holding costs and back-order penalty costs. We compare a traditional information
policy that does not use shared information with a full information policy that does exploit
shared information. In a numerical study we find that supply chain costs are 2.2% lower on
average with the full information policy than with the traditional information policy, and the
maximum difference is 12.1%. We also develop a simulation-based lower bound over all
feasible policies. The cost difference between the traditional information policy and the lower
bound is an upper bound on the value of information sharing: In the same study, that
difference is 3.4% on average, and no more than 13.8%. We contrast the value of information
sharing with two other benefits of information technology, faster and cheaper order
processing, which lead to shorter lead times and smaller batch sizes, respectively. In our
sample, cutting lead times nearly in half reduces costs by 21% on average, and cutting batches
in half reduces costs by 22% on average. For the settings we study, we conclude that
implementing information technology to accelerate and smooth the physical flow of goods
through a supply chain is significantly more valuable than using information technology to
expand the flow of information.
(Supply Chain; Multi-Echelon Inventory Management; Periodic Review Policies; Electronic Data
Interchange)

1. Introduction general belief within industry that capturing and

Information technology has had a substantial impact sharing real-time demand information is the key to
on supply chains. Scanners collect sales data at the improved supply chain performance. The purpose of

point-of-sale, and electronic data interchange (EDI) this research is to test this belief by rigorously mea-
allows these data to be shared immediately with all suring the value of information sharing and compar-
stages of the supply chain. The application of these ing this value to two other sources of supply chain
technologies, especially in the grocery industry, has improvement: reducing lead times and increasing
substantially lowered the time and cost to process an delivery frequency by reducing shipment batch sizes.
order, leading to impressive improvements in supply Note that the same information technology that facil-
chain performance (see Cachon and Fisher 1997, Clark itates information sharing also contributes to the re-
and Hammond 1997, Kurt Salmon Associates 1993). duction of lead times and shipment frequency by
As a result of these success stories, there is now a reducing the time and cost to process orders. Thus, the

MANAGEMENT SCIENCE ? 2000 INFORMS 0025-1909/00/4608/1032$05.00

Vol. 46, No. 8, August 2000 pp. 1032-1048 1526-5501 electronic ISSN

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

question addressed in here is not whether information nearly half (from five to three periods) reduces costs
technology improves supply chain performance, but by 21% on average, and cutting batch size in half
how. Specifically, does the primary gain come from reduces supply chain costs by 22%. We recognize that
sharing information or from allowing products to flow this comparison is meaningful only if those lead time
more quickly and evenly in the supply chain? and batch size reductions can be reasonably expected
We address this question within the context of a from the implementation of information technology.
supply chain with one supplier and N identical retail- In fact, we did observe comparable reductions at
ers that face stationary stochastic consumer demand Campbell Soup Company when it implemented infor-
with a known distribution. There are fixed transpor- mation technology to improve its supply chain.1 There
tation times between locations, and shipment quanti- has also been other documentation on the impact of
ties equal a multiple of a base batch quantity. There information technology in the grocery industry: Ba-

are holding costs at all levels and back-order penalty rilla, the world's largest pasta producer, reduced its

costs at the lowest level. This model provides a lead time from over one week to two days (Harvard

reasonable representation of supply chains selling an Business School case 9-694-046); and H.E.B., a large

established product under constant pricing condi- grocery chain based in Texas, eliminated 6 to 10 days

tions. from its lead time (Harvard Business School case

We consider two levels of information sharing. With 9-195-125). We conclude that while information shar-

traditional information sharing the supplier only ob- ing does reduce costs, simply flowing goods through
the supply chain more quickly and more evenly
serves the retailers' orders. With full information shar-
produces an order of magnitude greater improve-
ing the supplier has immediate access to the retailers'
ment.
inventory data. We develop an inventory policy for
The next section reviews the related literature. Sec-
each information sharing level. Reorder point policies
tion 3 outlines the model, and ?4 describes how to
are used with traditional information sharing. The
select inventory policies with traditional information
retailers also use reorder point policies with full
sharing. Section 5 details a lower bound over all
information, but the supplier does not. Instead, the
feasible policies. Section 6 develops the full informa-
supplier uses its additional information to better allo-
tion inventory policy. Section 7 details a numerical
cate inventory among the retailers and to improve its
study, ?8 discusses our results, and ?9 concludes.
order decisions (i.e., to better time its own replenish-
ments).
The difference between supply chain costs under
2. Literature Review
traditional and full information is one measure of the The following papers show how sharing demand and
value of shared information. However, there may exist inventory data can improve the supplier's order quan-
even better policies for either information level, that is, tity decisions in models with known and stationary
optimal policies are unknown for each level. To ac- retailer demand: Bourland et al. (1996), Chen (1998),
count for this possible bias, we develop a simulation-
based lower bound over all feasible policies, no matter
1 Campbell Soup Company gave us ordering data from several
what the level of information sharing is. The cost
retailers before and after the implementation of information tech-
difference between traditional information and the
nology. In the "before" data we saw that retailers often purchased in
lower bound is the maximum value of shared infor- multiple pallet quantities because, according to our contacts at
mation. Campbell Soup Company, they did not want to bother with the

In a numerical study with a wide range of param- hassle of placing orders frequently. In the "after" data it is clear that
each products' minimum batch size was no greater than one pallet,
eter values we find that information sharing reduces
and in some cases Campbell Soup Company was willing to deliver
supply chain costs by 2.2% on average, and the gap
in half-pallet increments. In addition, the lead time for deliveries to
between traditional information policy cost and the the retailers was reduced from about one week to two to three days,
lower bound is 3.4% on average. Cutting lead time by primarily resulting from the reduction in the order processing time.

MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000 1033

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

Gavirneni et al. (1999), and Aviv and Federgruen supply chain. Gavirneni et al. (1999) report that shar-
(1998). Lee et al. (2000) use shared information to ing the retailer's demand data reduced the supplier's
improve the supplier's order quantity decisions in a cost by 1%-35%.2 The impact on the supply chain's cost
serial system with a known autoregressive demand would be lower because information sharing in their
process. Liljenberg (1996) studies how to use shared model has no impact on the retailer's costs.
information to improve the supplier's allocation of There is other research related to our work. Lee et al.
inventory among the retailers. In our model shared (1997) find that sharing information reduces the sup-
information is exploited for both uses: better supplier plier's demand variance, which should benefit the
replenishments and better allocations to the retailers. supply chain, but they do not quantitatively measure
We focus on sharing demand and inventory data, this benefit. There are many studies that investigate a
but there are other data that can be shared in a supply supply chain model with one supplier, N retailers,

chain. Gavirneni et al. (1999) measure the benefit of stochastic consumer demand, and batch ordering.

sharing the parameters of the retailer's ordering policy Some of them assume traditional information (e.g.,

with the supplier. Aviv (1998) explores the benefits of Axsater 1993, Cachon 1995, Chen and Samroengraja

sharing forecasts for future demand. 1996, Lee and Moinzadeh 1986, Svoronos and Zipkin

In our model, as in the other studies mentioned, it is 1988), while others assume full information (e.g., Chen

assumed that information is always shared truthfully. and Zheng 1997, Graves 1996, McGavin et al. 1993).

Cachon and Lariviere (1997) study forecast sharing Because of different assumptions and test problems, it
is not possible to meaningfully compare supply chain
when the forecast provider has an incentive to provide
costs across those two sets of studies. Several research-
an overly optimistic forecast of demand.
ers study allocation rules, but none addresses the issue
Both Lee et al. (2000) and Gavirneni et al. (1999)
of information sharing (see Cachon 1995, Chen and
assume there exists a perfectly reliable exogenous
Samroengraja 1996, and Graves 1996). Anand and
source of inventory; information sharing has no im-
Mendelson (1997) study a one-period model in which
pact on the retailer because its orders are always
retailers possess some local information that cannot be
received in full after a fixed number of periods. In the
shared with either a central agent or other retailers. In
other papers, as in our model, the supplier is the only
our full information model all relevant information
source of inventory. Therefore, information sharing
can be shared with the central agent (i.e., the supplier).
may impact the retailers by changing the supplier's
Chen and Zheng (1994) develop a lower bound over
order quantities or allocations.
all feasible policies for a multiple retailer model. They
Gavirneni et al. (1999) and Aviv and Federgruen
show that a full information policy is reasonably close
(1998) allow for limited supplier capacity, whereas
to optimal, but they do not compare this policy with a
capacity is unrestricted in our model and in the other
traditional information policy. Graves (1996) also
papers.
shows that his full information policy is close to
The reported benefits of information sharing vary
optimal.
considerably. Liljenberg (1996) finds that better alloca-
tion lowers supply chain costs by 0% to 3.9%. Chen
(1998) finds that supply chain costs are lowered up to 3. Model
9%, and on average by 1.8%. Aviv and Federgruen Firm 0 is the supplier and firms [1, N] are identical
(1998) report benefits of 0%-5%. In contrast, Lee et al. retailers. The supplier replenishes its inventory from a
(2000) find that information sharing lowered supply perfectly reliable single source, that is, the supplier's
chain costs by about 23% in their scenario with the orders are always received after a constant lead time.
highest demand nonstationarity. However, Graves
(1999) studies a similar model, with the exception that
2 Those data are taken from the scenarios with the highest supplier
there is no outside inventory source, and concludes capacity. They found that the value of information sharing declined
that information sharing provides no benefit to the as the supplier's capacity decreased.

1034 MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

The supplier is the retailers' only source of inventory, route to the retailers, hs + hr per unit of stock at the
so there is no diversion of stock among the retailers retailers, where hr ' 0, and Pr per back order at each
and stockouts at the supplier cause replenishment retailer.
delays for the retailers. Inventory is reviewed period- Each shipment to a retailer equals an integer num-
ically, and within each period the following sequence ber of batches, where a batch is Qr units. Each supplier
of events occurs: (1) retailers order, (2) the supplier order is an integer multiple of QsQr units. Because all
orders, (3) inventory shipments are received and re- of the supplier's shipments to the retailers equal an
leased, (4) consumer demand occurs, and (5) inven- integer number of batches, all supplier variables are
tory holding and backorder costs are charged. measured in batches. When the supplier submits an
The following variables are defined before demand order in period t, it receives the entire order in period
t + LS. A batch shipped to a retailer in period t is
in period t (between Events 3 and 4) for each firm i
received in period t + Lr. The supplier will ship part
E [0, N]: Ii(t), on-hand inventory; Bi(t), outstanding
back orders; Oi(t), on-order inventory, but not yet of a multiple batch retailer order.
shipped; ITi(t), in-transit inventory; and IPi(t) = Ii(t)
- Bi(t) + ITi(t), inventory position.3 The supplier's
on-order inventory is always zero because its source 4. Traditional Information Policies
always ships inventory immediately, but a retailer's This section defines the inventory policies the firms
on-order inventory is positive when the supplier is use under traditional information. We then demon-
unable to fill the retailer's order completely. Variables strate how to evaluate expected supply chain cost for
without a period designation apply to period t, e.g., Iia given policy. A search over the set of reasonable
is retailer i's inventory in period t before consumer policies (defined later) yields the optimal policy.
demand.
4.1. Inventory Policies
A superscript "e" on a variable indicates that it is
The retailers implement a (Rr, nQr) reorder point
measured at the end of the period, and a superscript
policy: When Rr 2 IPb + ob, a retailer orders the
"b" indicates that it is measured at the beginning of
the period. (Naturally, IPe(t) = IPb(t + 1).) Dropping smallest integer multiple of Qr units to ensure that Rr
< IPr + or. The supplier implements an analogous
the subscript from a variable denotes the vector of that
variable across all firms, e.g., IP = tIPo, ... I IPNI. (Rs, nQs) reorder point policy. Reorder point policies
are probably not optimal in most cases, but the opti-
Since retailers are identical, it is convenient to associ-
mal policies under traditional information are not
ate a variable with a generic retailer: A subscript "r"
known. Nevertheless, reorder point policies are sim-
indicates a variable applies to any retailer, and a
ple to implement and intuitively reasonable. In fact,
subscript "s" identifies the supplier.
they are optimal in a serial supply chain with batch
Define D'T as demand at one retailer over T consec-
ordering (Chen 1997).
utive periods. Demand is discrete, independent, and
It remains to define the supplier's policy for allocat-
identically distributed across retailers and across time.
ing inventory among the retailers when the supplier's
We assume D E [0, d], d is a finite integer, and Pr(D 1
inventory is insufficient to cover the retailers' total
= 1) > 0. Let -tr = E[Dr. Demands not filled
orders in a period. We assume the supplier imple-
immediately from stock are backordered and eventu-
ments batch priority allocation. With this scheme the
ally filled.
supplier assigns in every period a priority to each
There are holding and back-order costs in each
batch ordered in that period. Suppose retailer i orders
period: h, > 0 per unit of stock at the supplier or en
b batches in a period. Then, the first batch in the order
is assigned priority b, the second batch is assigned
3 Some authors define inventory position to include on-order inven-
priority b - 1, and so forth. All batches ordered
tory. That definition is less cumbersome when working with tradi-
tional ordering policies but more cumbersome when working with within a period are placed in a shipment queue, and
the lower bound. they enter in decreasing priority order. If multiple

MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000 1035

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

batches have the same priority, a fair coin toss deter- its expected sojourn in inventory). Also using Little's
mines the sequence in which they enter the shipment Law and observing that with reorder point policies IP,
queue. Actual deliveries to the retailers are based on + OS and IPr + Or are uniformly distributed on the
the shipment queue on a first-in-first-out basis. There- intervals [Rs + 1, Rs + Qs] and [Rr + 1, Rr + Qr]'
fore, all batches ordered before period t are shipped respectively, it can be shown that
ahead of batches ordered in period t (but possibly in
the same period). Under batch priority the retailer
E[Br] = E[Ir] + I-Lr(E[U] + Lr + 1) -(Rr + 2 )
ordering the most batches within a period is presumed
to have the highest need, and therefore the first batch
je] Qs+ 1 NIJr
in its order is assigned the highest priority for the E[I] = RS + 2 + Q (E[U]-LS-1),
2 Qr
period. Of course, priorities only matter when inven-
tory is insufficient to cover the retailers' orders. where E[U] is the expected number of periods the
Even though batch priority is a reasonable alloca- supplier delays shipping a batch to a retailer. (See
tion rule given traditional information, there may exist Axsater 1997 for additional details.)
better allocation schemes. For example, because the It remains to evaluate E[U] and E[S]. Some ad-
supplier observes the time since each retailer's last ditional random variables are needed. When Rr ' IPpb
order, the supplier might use those data to better + Ob , define Vi(t) = R- IPib(t) _ Ob(t). Call Vi the
determine the retailers' priorities. However, we be- overshoot. From Cachon (1995),
lieve that it is computationally intractable to evaluate
expected costs under that allocation scheme. Further- Pr(Dlr? Qr + v) - Pr(Dr ?v)
Pr(VQ = v) = r-1 (1 - Pr(D1 ' c))
more, the difference between our traditional informa-
tion policy and the lower bound provides a conserva-
Let v) be the maximum overshoot, v = d - 1. Let br
tive estimate of the value of shared information.
be the number of batches a retailer orders because of

4.2. Evaluating Expected Costs an overshoot v,

Let E[C] be the expected per period supply chain
holding and back-order costs for a given pair of Iv I
reorder points, br(V) = j + 1.

E[C] = hsQrE [Ie] + N(hsE[IT e] Define UbV as the number of periods the supplier
delays shipping the bth batch in a retailer order
+ (hs + hr)E[I e] + prE[B]),
triggered by an overshoot v. From Cachon (1995),

where all of the above expectations depend on the

chosen reorder points. (They are omitted for nota- E[U]E b-l 'v= O E[Ubv] Pr(Vr = V)
[v= O Pr(Vr = v)br(V)
tional clarity.) The equations we provide evaluate
E[C] exactly when Rs ' -1, otherwise those same Taking a weighted average over all units in the supply
equations evaluate E[C] approximately. We begin chain yields
with several straightforward results. We then turn to
the evaluation of the supplier's shipping delay, i.e., the E[S]

number of periods the supplier delays shipping a

batch in a retailer's order. Evaluation of that distribu- Q b) C V=- Ili 1 E[Sbcvu] Pr(Vr = V) Pr(Ubv = U)
Qrb

tion function is the main challenge in the evaluation of Iv=O Pr(Vr = v)br(V
E[C].
(1)
From Little's Law, E[ITe] = KrLr and E[1e] = IrE[S],
where E[S] is the expected number of periods in where u- is the maximum shipping delay and Sbcvzi is
which a unit is charged holding costs at a retailer (i.e., the inventory sojourn of the cth unit in the bth batch of

1036 MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

an order that is triggered by an overshoot v and 4.3. The Supplier's Delay Distribution
experiences a shipping delay of u periods. (When R, Suppose in period t retailer j experiences an overshoot
- -1, iu = LS - 1, otherwise iu ' Ls - 1.) From v. Thus, Ubv is the number of periods the supplier
Cachon (1995), delays shipping the bth batch in retailer j's order. For
brevity, let batch b refer to that batch. Because the
00

supplier implements a reorder point policy, Cachon

E[SbCv,u= E Pr(Dr 'R,-v
T=U+Lr+l
(1995) shows that

+ (b1-)Qr + c-1). (2)

1 1;
When RS ' -1, (1) and (2) are exact, otherwise they
are approximations: When RS < -1, a retailer's lead
Pr(UbV ' u) =QS Pr(Ubvq ? u),
q=1

time demand is not independent of the lead time.

Now the only missing piece is the distribution func-
tion of U1,,. That distribution function is evaluated next. where

Pr(XB Ls-t < Rs + q-b) R + q-b 2 o, o C U < Ls

Pr(Ubvq C U) = 1 RS + q-b 2:O, u?2 Ls + 1
Pr(XFv-Ls-l > -1 - br(v) - (Rs + q - b)) RS

In the above, XB' is the number of batches the retailers + 1, t + T], and JB as the number of batches retailer
order over periods [t - , t], excluding batches in period j orders over periods [t - T, t - 1]. Define the
t shipped after batch b and all batches retailer j orders in following random variable for any retailer i,
period t. Similarly, XFT is the number of batches the
Wi(t) = Rr + Qr - IP(t) - Oi(t)
retailers order over periods [t, t + ], excluding batches
in period t shipped before batch b and all batches retailer Given that every Qr demand triggers a retailer order,
j orders in period t. Think of XB and XFT as the the (Wi) demand before the start of period t triggers

supplier's "time backward" and "time forward" demand retailer i's last order in period t or earlier. Also, the

processes relative to retailer j's period t order. Note that the (Qr - Wi)th subsequent demand after the start of
above results do not depend on the allocation scheme the period t triggers retailer i's first order to occur after

supplier implements. The evaluations of XB and XF T do period t. In general, let YT(Wi(t)) be the number of
depend on the allocation scheme. batches retailer i orders over periods [t + 1, t + Ti,
XB and XF are the summation of retailer j's order
process and the order processes of the other N - 1
YT(X) rL(X j
retailers. Refer to those other N - 1 retailers as the
"non-j" retailers. Hence, we begin with retailer j's
order process and then consider the order process of a Let w be the realization of Wj(t),
non-j retailer. Finally, these processes are combined to
yield XBV and XFV. Ivl
w = v - -jQr
4.3.1. Retailer j's Order Process. Define JFv as the
number of batches retailer j orders over periods [t It follows that JF = Y, (w).

MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000 1037

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

Let Wjv(t - 1) be Wj(t - 1) conditional on retailer YF0(Wk) = 0.5 min (b, Yr(Qr - 1 - Wk))
j's period t overshoot. From Bayes theorem,
+ 0.5 min (b - 1, Yr(Qr - 1 - Wk))-
Pr(Dr = ? Qr=- w)
Pr(Wjv(t - 1) = w) = Pr(V = v) Wk is independent of Vj, so Wk is also uniformly
distributed on the interval [0, Qr - 1]. Given Wk,
retailer k's order process in period t is independent of
Note that the (Wjv(t - 1) + 1) previous demand
its order process in periods [t + 1, t + T]. Hence,
triggered the last batch ordered by retailer j in period
t - 1 or earlier. Hence, JB = Y T(Q. - 1 -W
1 Qr-1 y
-1)).
Pr(YF<T y)r w=o
=-x=oE Pr(YT(w) = x)
4.3.2. The Non-j Retailers' Order Process. Con-
sider retailer k, where k / j. Define YBT as the number x Pr(YF?(w) ? y - x).
of batches retailer k orders over periods [t - , t],
4.3.3. The Order Process of N Retailers. Define
excluding any batches shipped after batch b. Define
NBT and NFT as the (N - 1)th fold convolutions of
YFT as the number of batches retailer k orders over
YBT and YFT, respectively;
periods [t, t + T], excluding any batches shipped
before batch b. Let X equal batch b's priority. XBvj = JBV + NB,
Begin with retailer k's period t order process. Re-
tailer k orders Yl(Wk(t - 1)) batches in period t. XFvj = JFV + NFT.
Because Wk(t - 1) is independent of Vj, Wk(t - 1) is
uniformly distributed on the interval [0, Qr - 1]. Because the ordering processes of retailers are inde-

According to batch priority, all batches retailer k pendent, the above are simple convolutions.
orders in period t with higher than b priority are
4.4. Finding Optimal Reorder Point Policies
shipped before batch b. (Recall that batch b's priority
For a fixed RS, E[C] is convex in Rr, so it is straig
is b.) Any batch with priority b is shipped before batch
forward to search for the optimal Rr given Rs. Ho
b with a 0.50 probability. Hence, let YB0(Wk(t - 1)) be
ever, E[C] is not necessarily jointly convex in Rs an
YB0 conditional on Wk(t - 1), R r* Hence, finding the optimal reorder points req
a search over feasible Rs values. Fortunately, it is
YB0(Wk(t - 1)) = 0.5(Yr(Wk(t - 1)) -
possible to constrain the search region. There is no
+ 0.5(Yr(Wk(t - 1)) - W + 1) - need to consider Rs < - Qs, because then increasing
Rs provides the retailers with a better lead time
For a given Wk(t - 1), retailer k's order process in
without increasing the supplier's inventory above
period t is independent of its order process in periods
zero. For a sufficiently large Rs there is no need to
[t - T, t - 1]. Hence,
consider an even larger Rs because then the supplier
(almost) always fills the retailers' orders immediately;
1 Qr-1 Y any further increase in Rs would increase the suppli-
Pr(YB T y =_ E E Pr( r (r - 1 - W) = X)
rw=O x=O
er's inventory without reducing costs at the retailers.
(See Cachon 1995 or Axsater 1997 for additional de-
X Pr(YB0(w) C y - x). tails.)

Now consider YFT. All batches with a priority lower

than b are certainly shipped after batch b, and a batch 5. Lower Bound
with priority b is shipped after batch b with 0.50 This section develops for this model a lower bound for
probability. Let YF(Wk) be YF conditional on Wk. supply chain costs that is independent of the level of
Because Yrl(Qr - 1 - Wk) is the number of batches information sharing. The first step divides the supply
retailer k orders in period t, chain's costs into two components. The second step

1038 MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

evaluates a lower bound for each component. The

final step sums the components' lower bounds to yield
a lower bound for the supply chain. Gp(y, R) = ;R Qr( L R&rY I
0 otherwise,
5.1. A Division of Supply Chain Costs
Actual supply chain holding and back-order costs in Gr(y, R) = Gr(Y) -Gp(y, R),
period t are
N

N G,(y, IP, R) = hs(y NILr) + E Gp(IPi, R),

C = hsQrIe + E [hsITe + (hs + hr)I + PrB]. (3)

and R* is the largest integer, such that Gr(R* + Qr)

Define the supplier's echelon inventory level, ILs, asGr(R*).

the
amount of inventory (in units) at the supplier or lower We now confirm that the expected sum of our

in the system minus total consumer backorders (be- charges equals expected actual costs. Because GP(IP
fore consumer demand in period t): R*) is merely a transfer payment between the supplier
and the retailers,
N
N
ILe e = e + [ITe + Je-Bl.
i=l1 GJ(ILs, IP, R*) + E Gr(IPi, R*) = hs(ILs - Nr)

ILS is measured in units (i.e., not in batches, as

N
are the
other supplier variables) because the supplier's eche-
+ E E[hrIe(t + Lr) + (hs+ Pr)B'(t + Lr)].
lon inventory level is not necessarily an integer mul- i=l

tiple of Q, units. Rewrite (3) as

Over an infinite horizon
N

C = h5ILse ? E [hrII + (hs + Pr)Bi. (4)

E[hrII(t + Lr) + (h,+ Pr)Be(t + Lr)]
= E[hrIt + (hs + Pr)Bij
Each period, the supply chain must decide an
inventory quantity to order from the supplier's source and within any period E[h,ILe] = E[h,(IL, -NP0]
and quantities to ship to the retailers. These shipment So
release decisions are made before the realization of
N
demand, yet the actual costs, (4), are incurred after
demand. To link the shipment release decisions to
E[C] = E[Gs(ILs, IP, R*)] + E E[Gr(IPi, R*)]. (5)
actual costs, we construct a system of charges that are
incurred before demand and equal actual costs in Thus, actual costs have been divided into two compo-
expectation. (We deliberatively use "charges" to refer nents: The first component is the supplier's charges
to our system of fees and use "costs" to refer to actual and the second component is the retailers' charges.
costs.)
5.2. Evaluation of the Lower Bound
In our system of charges, in period t retailer i is
We seek a lower bound for both the supplier's charges
charged Gr(IPi, R*) and the supplier is charged
and the retailers' charges. Begin with the retailers. Let
GJ(ILs, IP, R*), where
Cr = min E[Gr(IP)]. Since Gr(Y) is convex and all
retailer shipments must equal a multiple of Q r units, a
Gr(Y) = E[hr(Y - Dr)
reorder point policy minimizes expected costs. In fact,
+ (hs + Pr)(Y-D Lr+l) R* is the optimal reorder point: If IPi ' R* + 1,

MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000 1039

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

shipping another batch to the retailer raises its costs; if N

IPF R*, shipping another batch to the retailer lowers s.t. IPi<ILS,
its costs. When the supplier can raise IPi above R in
every period, under a reorder point policy IPi is IPi=IP'i+P~~iQr
uniformly distributed on the interval [R* + 1, R*
+ Qr] In that case, Each period the retailers' inventory positions are cho-
sen subject to two constraints: The sum of the retailers'
i Qr inventory positions cannot exceed the system's inven-
Cr = Qr Gr(R + c). tory level, IL,; and inventory can be added or sub-
Qr c=1
tracted from a retailer only in integer batch quantities.
The first constraint applies to any feasible policy. In
The reorder point R* also minimizes E[Gr(IPi, R*)], the second constraint, the free shuffling of inventory
even if the supplier is unable to raise IPi above R* inallows f3i< 0, whereas any feasible policy is restricted
every period: GP(IPi, R*) exactly compensates the
to 3i - 0. (Note that allowing a free shuffling of
retailer for any additional costs incurred due to the inventory among retailers is equivalent to allowing
supplier's less than perfect reliability. Furthermore, Cr negative orders.) Thus, C, + NCr is a lower bound f
= min E[Gr(IPi, R*)]. The function GP(IPi, R*) is supply chain costs, i.e.,
analogous to Clark and Scarf's (1960) induced penalty
function, because it equals the additional cost the min E[C] - Cs + NCr.
retailer incurs when the supplier fails to deliver im-
As with any feasible policy, a policy that yields Cs
mediately what the retailer requests. So NCr is a tight
must make two decisions: how inventory should be
lower bound for the retailers' charges; from (5), our
allocated among the retailers (what Pi to choose fo
lower bound is now
each retailer), and how much inventory the supplier
should order each period.
Begin with the allocation decision. Define H( y, r) as
min E[C] = min E[G,(IL5, IP, R*)] + NCr.
the change in penalty charges when an additional
batch is allocated to a retailer,
Turn to the supplier's charges. It is not known
which policy minimizes the supplier's charges be- H(y, r) = Gp(y, r) - Gp(y + Qrl r)
cause of the location constraint: Shipping a batch to
Note that H(y, R*) is nonincreasing in y. Therefore,
retailer i in period t might minimize expected costs in
penalty charges decline if IPF > IPi + Qr and one
period t, but does not necessarily minimize costs in batch is removed from retailer j and allocated to
future periods. To circumvent this problem, the loca- retailer i. So the first goal of the allocation decision is
tion constraint is relaxed, which means that batches
to balance the retailers' inventory positions: Move a
can be freely and instantly moved from one retailer to batch from the retailer with the highest inventory
another in each period. (This shuffling of batches position to the retailer with the lowest inventory
occurs before demand, i.e., when the shipment release position as long as the difference between their inven-
decisions are made.) Therefore, shipping a batch to a tory positions (before moving the batch) is greater
retailer with the highest need in period t does not than Qr. After balancing the retailers' inventory posi-
prevent the supply chain from moving that batch to a
tions the supplier allocates its inventory (whenever I's
higher need retailer in period t + 1. > 0, otherwise the supplier has no inventory to
Let Cs be expected supplier charges with the loca- allocate). The retailers' inventory positions are sorted
tion constraint relaxed, in ascending order and batches are allocated in that
sequence. Because Gp(y > R*, R*) = 0, the supplier
Cs = min E[Gs(ILs, IP, R*)] stops allocating inventory when either it runs out of

1040 MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

inventory or when the lowest retailer inventory posi- allocated to the retailer with the lowest inventory
tion is above R*. position. Let Zb(t + Ls) be that retailer's period t + LS
The supplier's ordering decision is more challeng- inventory position before this batch is allocated, that
ing than the allocation decision. We develop a myopic is, if that batch is allocated to that retailer, then that
policy and then argue that the myopic policy is indeed retailer's inventory position is increased from Zb(t
optimal. + Ls) to Zb(t + LS) + Qr. Thus, that batch will lower
Define the supplier's echelon inventory position, EI, period t + L, penalty charges by H(Zb(t + LS), R*).
= IL, + QrITs. Recall that ILs is measured in units and ITS Note that H(y > R*, R*) = 0 and H(y - Qr, R*)
is measured in batches, so QjnTs is the number of units in (h, + Pr)Qr Hence, to minimize period t + L,
transit to the supplier. Also define Gs(y, IP, R*), charges, the supplier should order the Oth set of Qs
batches if
Gs(y, IP, R*) = min h,(y - NDr9)
Is+ITb+ OQs
N
hsQrQs ' (hs + Pr)Qr E
+ j E Gp(IPi - D L + IiQr, R*) b=l+ITs+ (0 1)Q,+ 1

X Pr(Zb(t + LS) C-Qr)

N
R*
s.t. EIPi-D L, + f3iQr c yND L,
i=l
+ E Pr(Zb(t + Ls) = w)H(w, R*), (7)
W=-Qr+1

ie { . . ,-1, O, 1, . } .........(6)
where the right side above is the expected reduction in

In words, Gs(EIs, IP, R*) is the supplier's expected period t + L, penalty charges due to ordering those
charges in period t + Ls given that batches will be batches. Because H(y, R*) is decreasing in y and Z1+1
allocated among the retailers to balance their invento-stochastically dominates Z,, there exists a 0*'- 0 such
ries as much as possible. (Note that EIs - NDrLS is thethat (7) holds for all 0 ? 0*, but not for 0 > 0*. Thus,
supplier's echelon inventory level in period t + Ls.) given that the supplier has decided to order (0 - 1) Q,
The supplier's period t order is Els - EIP', where Els batches, it will reduce period t + L, charges if it orders
- EIs' must be a nonnegative integer multiple of QrQs. one more set of Qs batches as long as 0 ? 0*. In short,
From (6), each batch the supplier orders will certainly the myopic policy orders 0*Q, batches.
increase period t + LS costs by hsQr. However, each The implementation of (7) requires the distribution
batch the supplier orders will also reduce penalty function of Zb(t + Lj). Let Y#(IPj, R) equal the
charges. The supplier must order a multiple of QS minimum number of batches that must be allocated to

batches. Consider the Oth set of QS batches the sup- retailer i over periods [t + 1, t + Lj so that IPi(t
plier could order, 0 G {1, 2,.. .1. Assuming the sup- + LS) > R,
plier has ordered (0 - 1)Qs batches, the supplier
Pr(Yj(IPj, R) < x) =Pr(DLs <IPi-R + XQr- 1).
should order an additional Qs batches if the expected
reduction in penalty charges due to those batches is Let Y(IP, R) equal the minimum number of batches
larger than the expected increase in holding costs due that must be allocated to all retailers over periods [t
to those batches, h,QrQs. Hence, we need to evaluate + 1, t + Lj] so that they all have an inventory position
the expected reduction in penalty charges. greater than R,
In period t + LS the supplier will be able to allocate
N
to the retailers the batches it orders in period t plus the
Y(IP, R) = E Yi(IPi, R).
IS + IT' unallocated batches that are in the system in
period t. Consider the bth batch among those batches.
In period t + L, it might be allocated to some retailer. (Demand independence across retailers implies the
If it is allocated to a retailer, then it certainly will be above is a simple convolution.) If fewer than b batches

MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000 1041

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

are required to raise the inventory position of all of the Student t distribution with n - 1 degrees of freedom.4
retailers above R, then the bth batch is surely allocated Hence, confidence intervals can be constructed for C*.
to a retailer with an inventory position above R: Our estimate of the lower bound equals C* + NC_.
Chen and Zheng (1994) develop several lower
Pr(Zb(t + Ls) : R + 1) = Pr(Y(IP, R) c b - 1). bounds for multi-echelon inventory systems. Their
model is somewhat different: There are no order
We now argue that the myopic policy is optimal, quantity restrictions and they include explicit order-
ing costs. Nevertheless, they also divide the system's
i.e., yields mean cost C, Since Zb(t + LJ) stochasti-
costs into two components and evaluate lower bounds
cally dominates Zb(t + L, + 1), the Oth set of Qs
batches that the supplier could order in period t will for each one. However, since their model does not
include minimum order quantities, they are able to
reduce period t + L, + 1 penalty charges more than
explicitly evaluate each component's lower bound.
period t + L, penalty charges. So, if ordering the Oth
set of Qs batches is justified in period t (i.e., period t

+ L, penalty charges are reduced more than the

6. Full Information Policy
additional holding costs those batches create), then
Full information provides the supplier with data to:
ordering them is certainly justified in any later period.
(1) improve its order quantity decisions and (2) to
In other words, the actions taken in period t to
improve its allocation decisions. It can make better
minimize costs in period t + L, do not prevent the the supplier's local inventory data are
orders because
minimization of costs in subsequent periods. Because not a perfect proxy for the supply chain's replenish-
the myopic policy in period t does not interfere with ment need: They do not measure the supply chain's
the myopic policy in future periods, the myopic policy total inventory nor do they capture how the supply
minimizes costs in every period, i.e., it is optimal. chain's inventory is allocated among the retailers.
Although the described policy yields E[C3] over an Allocation of inventory can be enhanced for two
infinite horizon, there unfortunately does not appear reasons: (1) allocations can be based on the retailers'
to be a simple method to evaluate E [C] exactly. (The inventory positions rather than the number of batches
state space for IP is quite large, and it is not known they order (when Qr is large each retailer might order
how to evaluate the steady state distribution of IP one batch, but their inventory positions may differ
under the myopic policy.) Therefore, as a practical substantially); and (2) the allocation of a batch can be

solution we use simulation to estimate based

E[CJ]. We on the retailers' inventory positions in the
initialize the simulation with IPb = 0 and IPi' equal to period the batch is shipped rather than in the period
a random draw from a uniform random variable over the batch is ordered (as is done with batch priority).

the interval [- Q, 0], i.e., the system begins with no Although the lower bound decision rule (7) is not
optimal if batches cannot be freely moved among
inventory. We begin collecting cost data L, periods
after the first supplier order and stop collecting cost retailers, it does do, at least in part, what a full

data L3 - 1 periods after the yth supplier order, for information policy should do: Its order recommenda-

some y - 1. Let c' be the mean cost in the jth tion is influenced by both the total amount of inven-
tory in the system as well as how that inventory is
simulation. After n simulations, let C* be the estimate
distributed within the system. Since that rule is rea-
of E[Cs],
sonable, we assume that under full information the

1 n

S- =- c 4 For each scenario we tested in the numerical study we evaluated

j=1
the kurtosis of the data. (The kurtosis of a normal random variable
is zero). We constructed a QQ plot for the 10 scenarios with the
Assuming cg are drawn from identical and indepen- highest absolute kurtosis values, which revealed that even for those
dent normally distributed random variables, C* has a scenarios it is reasonable to assume the normal distribution.

1042 MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

supplier uses (7) to decide its order quantity each information policy will perform better than the tradi-
period. tional information policy. However, the numerical
We conjecture that (7) will tend to order fewer study finds that it did perform better in almost all of
batches than optimal because it underestimates the the tested scenarios.
value of inventory for reducing penalty charges: The
expected inventory position of the retailer allocated
7. Numerical Study
the bth batch in the supplier's order will be lower
This section reports on a numerical study that evalu-
without the free shuffling of inventory than with the
ates supply chain costs with the traditional informa-
free shuffling of inventory. However, this bias occurs
tion policy, the full information policy, and the lower
only if the retailers' inventory positions are out of
bound. Section 7.1 details the parameters and methods
balance, i.e., only if a free shuffling is useful. We
and ?7.2 details the results.
conjecture that out of balance situations are less likely
as Qr is increased, consumer demand variability is 7.1. Parameters and Methods

decreased, or as Ls is decreased. The numerical study consists of the 768 scenarios

The full information policy must also specify how formed from all combinations of the following param-
inventory is allocated/shipped to the retailers. The eters:

lower bound policy uses a retailer reorder point, R*, to

N 1{4, 16} Qr E {2, 4, 8, 16} Lr {1, 5}
decide when to ship inventory to the retailers, but it is
easy to confirm with a few numerical examples that R* hr 1 -Ihs Pr E {5, 25} crr E {0.36, 1}
is generally a poor choice in actual implementation.5
Unfortunately, there does not appear to be a simple Qs E {1, 4, 16} Ls E {1, 5} hs E {0.5, 1}.

technique for choosing the best reorder point. There- In all scenarios / r = 1. The parameters are chosen
fore, we implement the retailer's optimal reorder point reflect a wide range of situations: small and large
under traditional information; that reorder point is batch sizes (relative to mean demand), short and long
substituted for R* in (7). Note that this allocation lead times, cheap and expensive supplier inventory,
policy is an improvement over batch allocation because low and high back-order penalty costs, and low and
the supplier will always ship a batch to the retailer high consumer demand variability. When cRr = 1,
with the lowest inventory position in the period the demand at each retailer is Poisson. The distribution is
batch is shipped rather than to a retailer that had the truncated, so that Pr(D' c 7) = 1, whereas in the
lowest inventory position in some prior period (when actual distribution Pr(Di c 7) = 0.99999.6 When or
the batch was ordered). - 0.36, demand is a discretization of a Normal
As with the lower bound, a priori evaluation of distribution with mean 1 and standard deviation 3:
expected per period costs with this full information
policy is not possible. Hence, we evaluate this policy Pr(D' = 0) = 0.066807; Pr(D' = 1) = 0.866386;
using the same simulation approach used to evaluate
Pr(D' = 2) = 0.066807; Pr(D1 3) = 0.
the lower bound. There is no guarantee that this full
Although we do not expect to observe the above

5 Intuition also suggests that R* is not a good choice. When hr is distribution in practice, it provides a representation of
small relative to h5, R* is quite large. That means that once supply chain behavior with low consumer demand
inventory arrives at the supplier it will be immediately shipped to variability.
the retailers. (If the supplier held the inventory, then it would fill the For each scenario the system-wide cost-minimizing
retailers' orders quickly and their average inventory would be quite
reorder points are found assuming traditional infor-
high.) Pushing inventory down to the retailers immediately is not
risky when inventory can be freely moved among the retailers, but mation. The costs evaluations are exact when RS
it can be quite costly if rebalancing is not possible. At the other
extreme, if hr is large relative to h5 then R* will be low and the 6 The mean of that distribution is 0.99999 and the standard deviation
retailers will probably run higher back orders than they should. is 0.99992.

MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000 1043

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

Figure 1 Traditional vs. Full Information

500
Mean = 2.2%

450 Median = 1.8%

400

350

i 300

i; ~~~~~~~~249
0 250 24

200 173

150 13

100 69
51

50 36

0 1 3 212 8 2 1 3 0 1 0
0

-3% -2% -1% 0% 1% 2% 3% 4% 5% 6% 7% 8% 9% 10% 11% 12% 13% 14%

(Traditional Information Cost - Full Information Cost) / Traditional Information Cost

--1, and approximate otherwise. We confirmed via information policy, i.e., our realized value of information
simulation that the approximation introduces little sharing. Those differences, as well as the cost differences
error. displayed in the other figures, are all expressed as
To estimate the full information policy cost and the percentages of the traditional information policy cost.
lower bound we conducted 40 simulations of each The mean benefit of the full information policy is 2.2%,
scenario, where each scenario collected data over 10 the maximum benefit is 12.1%, and for 95% of scenarios
supplier order cycles (y = 10). Due to sampling error, the percentage is no more than 5.6%.
it is possible to observe a mean lower bound that is Figure 2 displays a histogram of the cost difference
greater than the mean full information policy cost. For between the full information policy and the lower
those scenarios in which that occurred, we disposed of bound, i.e., the potential cost improvement if the
the original data and conducted 90 additional simula- optimal full information policy were implemented.
tions per scenario. These latter data are used in our The mean is 1.2%, the maximum is 10.5%, and for 95%
analysis.7 For each scenario we constructed 95% con- of scenarios the percentage is no more than 4.1%. (In a
fidence intervals for both the lower bound and the full few scenarios the lower bound estimate exceeds the
information policy cost. The mean confidence interval, feasible policy estimate, which we attribute to sam-
expressed as a percentage of the mean estimate, is pling error.) We conclude that the performance of our
0.5% for the lower bound and 0.8% for the full full information policy is quite close to optimal.
information policy cost. We conclude that the esti- Figure 3 presents the upper bound on the value of
mates are sufficiently accurate. shared information: the cost difference between the
traditional information policy and the lower bound.
7.2. Results
Figure 1 displays a histogram of the cost difference
The mean value is 3.4%, the maximum is 13.8%, and
for 95% of scenarios the value is no more than 8.0%.
between the traditional information policy and the full
We investigated two other possible benefits of infor-
mation technology. One is that information technology
7 Due to this procedure there is a positive bias for the difference
between the full information cost and the lower bound cost in the reduces order processing times, which we model as a
scenarios that are not discarded. reduction in Lr. The second benefit is that information

1044 MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

Figure 2 Full Information vs Lower Bound

500 -
Mean = 1.2%

450- 422
450 ~~~~~422
Median = 0.6%

400

350

5 300 -

O 250

E~ 200||l

150 - 123

100 1 82

50 39~
0 3 316 13 8 4 0 2 1 0 0 0
-3% -2% -1% 0% 1% 2% 3% 4% 5% 6% 7% 8% 9% 10% 11% 12% 13% 14%

(Full Information Cost - Lower Bound) / Traditional Information Cost

Figure 3 Maximum Value of Shared Information

500
Mean 3.4%

450 - Median 2.8%

400

350

i 300

O 250

-D 200 179

z 150 160

102
100 85
67 69

50 3

0 0 0 0 ~~~921 16 14 5 2
-3% -2% -1% 0% 1% 2% 3% 4% 5% 6% 7% 8% 9% 10% 11% 12% 13% 14%

(Traditional Information Cost - Lower Bound) / Traditional Infofnation Cost

technology reduces order processing costs. We model tion, we feel that these are reasonable changes to the Lr
and Qr parameters.) In each scenario we assume tradi-
that as a reduction in Qr, because a firm can justify more
frequent ordering with lower order processing costs. tional information sharing. Table 1 displays data on the
Although the magnitude of these effects will depend on potential cost reductions. Across all scenarios it is appar-
the context, we consider the impact of cutting batch sizes ent that a reduction in Lr or Qr can have a significant
in half or the impact of reductions in Lr from 5 to 3, from impact on supply chain costs and a significantly greater
3 to 1, and from 5 to 1. (As we mention in the introduc- impact than sharing information.

MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000 1045

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

Table 1 Shared Information vs. Lead Time Reduction or Batch Size Reduction

% Decrease in Total Supply Chain Cost

Retailer Lead Time Reduction Retailer Batch Size Reduction

With Traditional Information With Traditional Information
Upper Bound on the Value
of Complete Information* 3 to 1 5 to 3 5 to 1 16 to 8 8 to 4 4 to 2 2 to 1

Minimum 0% 5% 4% 9% 16% 8% 3% 1%
Mean 3% 29% 21% 43% 36% 27% 17% 10%
Median 3% 30% 22% 46% 36% 26% 15% 7%
Maximum 14% 57% 35% 71% 49% 47% 43% 37%

* (Traditional Information Cost - Lower Bound)/Tra

8. Discussion of information technology on lead time and batch size

Our results are surprising. Indeed, we undertook than in facilitating information sharing. Further, be-

this research with the strong expectation that we cause we measure the value of information sharing
would be able to demonstrate significant benefits to using a lower bound on any feasible inventory policy
information sharing in these models. So why do the that uses shared information, this finding will not be
data speak the opposite conclusion? We conjecture changed by the formulation of more sophisticated
the following answer: The retailers' orders convey algorithms for sharing information.
to the supplier a substantial portion of the informa- Although our model is representative of many
tion the supplier needs to perform its ordering and actual supply chains, we recognize that our conclusion
allocation functions. When a retailer is flush with is limited to the setting we consider. In particular, in
inventory, its demand information provides little our model demand is known, the retailers are identi-
value to the supplier because the retailer has no cal, there exists only a single source for inventory,
short-term need for an additional batch. A retailer's there are no capacity constraints, there are no incen-
demand information is most valuable when the tive conflicts among the supply chain's firms, and
retailer's inventory approaches a level that should firms choose rational ordering policies.
trigger the supplier to order additional inventory, We anticipate that information sharing can have a
but this is also precisely when the retailer is likely to significantly greater value in environments with un-
submit an order. Hence, just as the retailer's de- known demand, for example, early sales of new
mand information becomes most valuable to the products or established products on promotion. In
supplier, the retailer is likely to submit an order, those settings information sharing would improve the
thereby conveying the necessary information with- supplier's ability to detect shifts in the demand pro-
out explicitly sharing demand data. cess.
How can we reconcile our results with the clear We study a model with identical retailers for two
trend across many industries to apply information reasons. First, nonidentical retailer models are far
technology to logistics and inventory management? In more complex to analyze. With nonidentical retail-
fact, our results are quite consistent with that trend: ers it is not even known how to evaluate reorder
We do find substantial savings from lead time and point policies with traditional information. Second,
batch size reductions, both of which are facilitated by in our setting we anticipate that information has the
the implementation of information technology. We highest value with identical retailers. Information
only conclude that the observed benefits of informa- sharing allows the supplier to identify which retail-
tion technology in practice are due more to the impact ers have the highest need for replenishments. This is

1046 MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

most important when the retailers are nondistin- Aviv, Y. 1998. The effect of forecasting capability on supply chain
coordination. Working paper, Washington University, St.
guishable, i.e., when they have identical demand
Louis, MO.
processes. In fact, Aviv and Federgruen (1998) ob-
, A. Federgruen. 1998. The operational benefits of information
tain comparable results to ours in a model with
sharing and vendor managed inventory (VMI) programs.
nonidentical retailers. Our justifications notwith- Working paper, Washington University, St. Louis, MO.
standing, additional research is needed to assess Axsater, S. 1993. Exact and approximate evaluation of batch-order-

information sharing with nonidentical retailers. ing policies for two-level inventory systems. Oper. Res. 41
777-785.
Although a single inventory source is reasonable for
. 1997. Exact analysis of continuous review (R, Q)-policies in
many settings, in some supply chains the firms may
two-echelon inventory systems with compound Poisson de-
have access to a second inventory source, albeit at a
mand. Working paper, Lund University, Lund, Sweden.
higher cost. In those settings information sharing may Bourland, K., S. Powell, D. Pyke. 1996. Exploiting timely demand
allow the supply chain to better decide when it should information to reduce inventories. Eutropean J. Oper. Res. 9
utilize its alternative sources. 239-253.

Gavirneni et al. (1999) found that information Cachon, G. 1995. Exact evaluation of batch-ordering policies in
two-echelon supply chains with periodic review. Oper. Res.
sharing is most valuable when capacity is not re-
Forthcoming.
strictive; information is valuable only if the system
, M. Fisher. 1997. Campbell Soup's continuous product replen-
has the flexibility to respond to the information. ishment program: Evaluation and enhanced decision rules.
Hence, imposing a capacity constraint on the sup- Prod. Oper. Management 6 266-276.

plier would probably lower the value of information , M. Lariviere. 1997. Contracting to assure supply: How to

in our model. share forecasts in a supply chain. Working paper, Kellogg

Graduate School of Management Northwestern University,
We have assumed that a benevolent dictator de-
Evanston, IL.
cides all inventory shipments. This is reasonable
Chen, F. 1997. Optimal policies for multi-echelon inventory prob-
when the sole objective is minimizing total supply lems with batch ordering. Working paper, Columbia Univer-
chain cost, in other words, it doesn't matter which sity, New York.
firm makes the decision. However, in actual supply . 1998. Echelon reorder points, installation reorder points, and

chains the firms might not share the same objectives. the value of centralized demand information. Management Sci.
44 S221-S234.
Additional research is needed to determine how the
, R. Samroengraja. 2000. A staggered ordering policy for one-
firms will behave in those settings. For example,
warehouse multi-retailer systems. Oper. Res. 48(2) 281-293.
they may needlessly hoard inventory, thereby rais- , Y. S. Zheng. 1994. Lower bounds for multi-echelon stochastic
ing costs for everyone. It is important to determine inventory systems. Management Sci. 40 1426-1443.

which conditions provide the players with the in- , ~~. 1997. One-Warehouse multi-retailer systems with cen-

centive to truthfully reveal their private information tralized stock information. Oper. Res. 45 275-287.
Clark, A. J., H. E. Scarf. 1960. Optimal policies for a multi-echelon
and whether revealing information eliminates de-
inventory problem. Management Sci. 6 475-490.
structive gaming.8
Clark, T., J. Hammond. 1997. Reengineering channel reordering
processes to improve total supply chain performance. Prod.
Oper. Management 6 248-265.
For many helpful comments, the authors thank Lee Schwarz,
Graves, S. 1996. A multiechelon inventory model with fixed replen-
Sridhar Tayur, Paul Zipkin, and seminar participants at the Depart-
ishment intervals. Management Sci. 42 1-18.
ment of Industrial Engineering and Engineering Management at
. 1999. A single-item inventory model for a nonstationary
Stanford University and the Operations Research Center at the
demand process. Manufacturing Ser. Oper. Management 1
Massachusetts Institute of Technology. The assistance of the asso-
50-61.
ciate editor and the reviewers is graciously acknowledged.
Gavirneni, S., R. Kapuscinski, S. Tayur. 1999. Value of information
in capacitated supply chains. Management Sci. 45 16-24.
References Kurt Salmon Associates Inc. 1993. Efficient Consumer Response:
Anand, K., H. Mendelson. 1997. Information and organization for Enhancing Consumer Value in the Grocery Industry. Food
horizontal multimarket coordination. Management Sci. 43 1609- Marketing Institute, Washington, DC.

1627. Lee, H., K. Moinzadeh. 1986. Batch size and stocking levels in

MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000 1047

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms
 CACHON AND FISHER
Supply Chain Inventory Management

multi-echelon repairable systems. Management Sci. 32 1567- two-echelon inventory system with stochastic demand. Work-
1581. ing paper, Lund University, Lund, Sweden.
, P. Padmanabhan, S. Whang. 1997a. Information distortion in a McGavin, E., L. Schwarz, J. Ward. 1993. Two-Interval inventory
supply chain: The bullwhip effect. Management Sci. 43 546-558. allocation policies in a one-warehouse N-identical retailer dis-
, K. So, C. Tang. 2000. The value of information sharing in a tribution system. Management Sci. 39 1092-1107.
two-level supply chain. Management Sci. 46(5) 626-643. Svoronos, A., P. Zipkin. 1988. Estimating the performance of multi-
Liljenberg, Peter. 1996. The value of centralized information in a level inventory systems. Oper. Res. 36 57-72.

Accepted by Luke Van Wassenhove; received March 13, 1997. This paper was with the authors 12 months for 2 revisions.

1048 MANAGEMENT SCIENCE/Vol. 46, No. 8, August 2000

This content downloaded from 45.5.165.8 on Thu, 22 Aug 2019 15:37:22 UTC
All use subject to https://about.jstor.org/terms