Inventory Management of A Fast-Fashion Retail Network: Felipe Caro Jérémie Gallien

informs
OPERATIONS RESEARCH
Vol. 58, No. 2, MarchApril 2010, pp. 257273
issn 0030-364X eissn 1526-5463 10 5802 0257
doi 10.1287/opre.1090.0698
2010 INFORMS
OR PRACTICE
Inventory Management of a Fast-Fashion

Retail Network
Felipe Caro
Anderson School of Management, University of California at Los Angeles, Los Angeles,
California 90095, [email protected]
Jrmie Gallien
Sloan School of Management, Massachusetts Institute of Technology, Cambridge, Massachusetts 02142,
[email protected]
Working in collaboration with Spain-based retailer Zara, we address the problem of distributing, over time, a limited amount
of inventory across all the stores in a fast-fashion retail network. Challenges specic to that environment include very short
product life cycles, and store policies whereby an article is removed from display whenever one of its key sizes stocks
out. To solve this problem, we rst formulate and analyze a stochastic model predicting the sales of an article in a single
store during a replenishment period as a function of demand forecasts, the inventory of each size initially available, and the
store inventory management policy just stated. We then formulate a mixed-integer program embedding a piecewise-linear
approximation of the rst model applied to every store in the network, allowing us to compute store shipment quantities
maximizing overall predicted sales, subject to inventory availability and other constraints. We report the implementation of
this optimization model by Zara to support its inventory distribution process, and the ensuing controlled pilot experiment
performed to assess the models impact relative to the prior procedure used to determine weekly shipment quantities. The
results of that experiment suggest that the new allocation process increases sales by 3% to 4%, which is equivalent to
$275 M in additional revenues for 2007, reduces transshipments, and increases the proportion of time that Zaras products
spend on display within their life cycle. Zara is currently using this process for all of its products worldwide.
Subject classications: industries: textiles/apparel; information systems: decision-support systems; inventory/production:
applications, approximations, heuristics.
Area of review: OR Practice.
History: Received November 2007; revisions received May 2008, September 2008, November 2008; accepted November
2008. Published online in Articles in Advance August 12, 2009.
1. Introduction
only 2,0004,000 items for key competitors. This increases

Zaras appeal to customers: A top Zara executive quoted in
Fraiman et al. (2002) states that Zara customers visit the
store 17 times on average per year, compared to 3 to 4 visits
per year for competing (non-fast-fashion) chains. In addition, products offered by fast-fashion retailers may result
from design changes decided upon as a response to actual
sales information during the season, which considerably
eases the matching of supply with demand: Ghemawat and
Nueno (2003) report that only 15%20% of Zaras sales are
typically generated at marked-down prices compared with
30%40% for most of its European peers, with an average
percentage discount estimated at roughly half of the 30%
average for competing European apparel retailers.
The fast-fashion retail model just described gives rise
to several important and novel operational challenges. The
work described here, which has been conducted in collaboration with Zara, addresses the particular problem of
distributing, over time, a limited amount of merchandise
inventory between all the stores in a retail network. Note
The recent impressive nancial performance of the Spanish

group Inditex (its 2007 income-to-sales ratio of 13.3% was
among the highest in the retail industry) shows the promise
of the fast-fashion model adopted by its agship brand
Zara; other fast-fashion retailers include Sweden-based
H&M, Japan-based World Co., and Spain-based Mango.
The key dening feature of this new retail model lies
in novel product development processes and supply chain
architectures relying more heavily on local cutting, dyeing,
and/or sewing, in contrast with the traditional outsourcing
of these activities to developing countries. Although such
local production obviously increases labor costs, it also provides greater supply exibility and market responsiveness.
Indeed, fast-fashion retailers offer in each season a larger
number of articles produced in smaller series, continuously
changing the assortment of products displayed in their
stores: Ghemawat and Nueno (2003) report that Zara offers
on average 11,000 articles in a given season, compared to
257
Caro and Gallien: Inventory Management of a Fast-Fashion Retail Network
258
Figure 1.
Operations Research 58(2), pp. 257273, 2010 INFORMS
Legacy process and new process envisioned to determine weekly shipments to stores.
(a) Legacy process
(b) New process envisioned
Assortment decisions
(the offer)
Store inventory
Assortment decisions
(the offer)
Requested shipment
Past sales data
quantities for each
Past sales data
reference and size
Forecasting
model
Store
managers
Inventory
in stores,
past sales
Requested shipment quantities

for each reference and size
Warehouse
inventory
Inventory
in stores
Demand
forecasts
Warehouse
allocation team
Optimization
model
Shipments
Shipments
that although the general problem just stated is not specic

to fast-fashion retailing, we believe that several features
that are specic to this retail paradigm (short product life
cycles, unique store inventory display policies) justify new
approaches. Indeed, Zaras interest in this area of collaboration was motivated by its desire to improve the inventory
distribution process it was using at the beginning of our
interaction for deciding the quantity of each article to be
included in the weekly shipment from the warehouse to
each store (see Figure 1(a) for an illustration).
According to that process, which we call the legacy process, each store manager would receive a weekly statement
of the subset of articles available in the central warehouse for which he/she may request a shipment to his/her
store. Note that this weekly statement (dubbed the offer)
would thus effectively implement any high-level assortment
decision made by Zaras headquarters for that particular
store. However, it would not mention the total quantity
of inventory available in the warehouse for each article
listed. After considering the inventory remaining in their
respective stores, store managers would then transmit back
requested shipment quantities (possibly zero) for every size
of every one of those articles. A team of employees at the
warehouse would then reconcile all those requests by modifying (typically lowering) these requested shipment quantities so that the overall quantity shipped for each article
and size was feasible in light of the remaining warehouse
inventory.
At the beginning of our interaction, Zara expressed some
concerns about the process just described, stating that
although it had worked well for the distribution network
for which it had originally been designed, the growth of its
network to more than a thousand stores (and recent expansion at a pace of more than a hundred stores per year)
might justify a more scalable process. A rst issue centered
on the incentives of store managers, whose compensation
Warehouse
inventory
and career promotion prospects are driven to a signicant

degree by the total sales achieved in their stores. We believe
that as a consequence, store managers frequently requested
quantities exceeding their true needs, particularly when suspecting that the warehouse may not hold enough inventory of a top-selling article to satisfy all stores (among
others, Cachon and Lariviere 1999 study a stock-rationing
model capturing this behavior). Another issue is that store
managers are responsible for a large set of tasks beyond
determining shipment quantities, including building, sustaining, and managing a team of several dozen sales associates in environments with high employee turnover, and
are thus subject to important time pressures. Finally, we
also believe that the very large amount of data that the
warehouse allocation team was responsible for reviewing
(shipments of several hundred articles offered in several
sizes to more than a thousand stores) also created significant time pressures that made it challenging to balance
inventory allocations across stores and articles in a way that
would globally maximize sales. Motivated by these observations, we started discussing with Zara the alternative new
process for determining these weekly shipment quantities,
which is illustrated in Figure 1(b). The new process consists of using the shipment requests from store managers
along with past historical sales to build demand forecasts. It
then uses these forecasts, the inventory of each article and
size remaining both in the warehouse and each store, and
the assortment decisions as inputs to an optimization model
having shipment quantities as its main decision variables.
The forecasting model considered takes as input from
store managers their shipment requests, which is the very
input they provide in the legacy process. This approach
was believed to constitute the easiest implementation path,
because it does not require any changes in the communication infrastructure with the stores or the store managers incentives. Note that Zaras inventory distribution
259
Table 1.
Main features of representative periodic review, stochastic demand models for inventory management in
distribution networks.
Decision scope
Time horizon
Shortage model
Retailers
Lost
NonPull back
Ordering Withdrawal Allocation Finite Innite Backorder sales Identical identical display policy
Eppen and Schrage (1981)
Federgruen and Zipkin (1984)
Jackson (1988)
McGavin et al. (1993)
Graves (1996)
Axster et al. (2002)
This paper
process could be further improved in the future by introducing explicit incentives for the stores to contribute accurate forecasts.1 However, the implementation reported here
shows that substantial benets can be obtained without any
such change in the incentive structure.
Although the forecasting component of the new process provides a critical input, we also observe that the
associated forecasting problem is a relatively classical one.
In addition, the forecasting and optimization models supporting this new distribution process are relatively independent from each other, in that both may be developed and
subsequently improved in a modular fashion. For these reasons, and for the sake of brevity and focus, the remainder
of this paper is centered on the optimization component,
and we refer the reader to Correa (2007) for more details
and discussion on the forecasting model developed as part
of this collaboration.
We proceed as follows: After a discussion of the relevant literature in 2, we discuss in 3 the successive steps
we followed to develop the optimization model, specically
the analysis of a single-store stochastic model (3.1) and
then the extension to the entire network (3.2). Section 4
discusses a pilot implementation study we conducted with
Zara to assess the impact of our proposed inventory allocation process. Finally, we offer concluding remarks in 5.
The online appendix contains a technical proof, a validation of the store inventory display policy, a detailed computation of the nancial impact, a model extension that
considers articles offered in multiple colors, and some additional material related to the software implementation of
this work. An electronic companion to this paper is available as part of the online version that can be found at http://
or.journal.informs.org/.
2. Literature Review
The fast-fashion retail paradigm described in the previous section gives rise to many novel and interesting operational challenges, as highlighted in the case studies on
Zara by Ghemawat and Nueno (2003) and Fraiman et al.
(2002). However, we are aware of only one paper besides
the present one describing an analytical model formulated
to address an operational problem that is specic to fastfashion companies. Namely, Caro and Gallien (2007) study
the problem of dynamically optimizing the assortment of
a store (i.e. which products it carries) as more information
becomes available during the selling season. The present
paper constitutes a logical continuation to that previous
work because Zaras inventory allocation problem takes the
product assortment as an exogenous input (see Figure 1).
The generic problem of allocating inventory from a
central warehouse to several locations satisfying separate
demand streams has received much attention in the literature. Nevertheless, the optimal allotment is still an open
question for most distribution systems. When demand is
assumed to be deterministic however, there are very effective heuristics with data-independent worst-case performance bounds for setting reorder intervals (see Muckstadt
and Roundy 1993 for a survey). For the arguably more
realistic case of stochastic demand that we consider
here, available performance bounds depend on problem
data. Focusing on stochastic periodic-review models (Zara
replenishes stores on a xed weekly schedule), Table 1
summarizes the main features of representative existing
studies along with that of the present one. A rst feature is the scope of inventory decisions considered: ordering refers to the replenishment of the warehouse from an
upstream retailer; withdrawal to the quantity (and sometimes timing) of inventory transfers between the warehouse
and the store network; and allocation to the split of any
inventory withdrawn from the warehouse between individual stores. For a more exhaustive description of this body
of literature, see Axster et al. (2002) or the earlier survey
by Federgruen (1993).
We observe that the operational strategy of fast-fashion
retailers consists of offering through the selling season a
large number of different articles, each having a relatively
short life cycle of only a few weeks. As a rst consequence, the innite-horizon timeline assumed in some of
the papers mentioned above does not seem appropriate
here. Furthermore, typically at Zara a single manufacturing order is placed for each article, and that order tends to
be fullled as a single delivery to the warehouse without
subsequent replenishment. Ordering on one hand and withdrawal/allocation on the other thus occur at different times,
260
and in fact, Zara uses separate organizational processes for
these tasks. Consequently, we have chosen to not consider
the ordering decisions and assume instead that the inventory available at the warehouse is an exogenous input (see
Figure 1). Although we do consider the withdrawal decisions, it should be noted that these critically depend in our
model on an exogenous input by the user of a valuation
associated with warehouse inventory, and any development
of a rigorous methodology for determining the value of that
parameter is beyond the scope of this work (see 3.2 for
more details and discussion). We also point out that Zara
stores do not take orders from their customers for merchandise not held in inventory, which seems to be part of a
deliberate strategy (Fraiman et al. 2002). This justies the
lost-sales model we consider.
The most salient difference between our analysis and
the existing literature on inventory allocation in distribution
networks is arguably that our model, which is tailored to
the apparel retail industry, explicitly captures some dependencies across different sizes and colors of the same article.
Specically, in Zara stores (and we believe many other
fast-fashion retail stores) a stockout of some selected key
sizes or colors of a given article triggers the removal (or
pull back) from display of the entire set of sizes or colors.
While we refer the reader to 3.1.1 and E (in the online
appendix) for a more complete description and discussion
of the associated rationale, that policy effectively strikes
a balance between generating sales on one hand, and on
the other hand mitigating the shelf-space opportunity costs
and negative customer experience associated with incomplete sets of sizes or colors. The literature we have found
on these phenomena is scarce, but consistently supports the
rationale just described: Zhang and Fitzsimons (1999) provide evidence showing that customers are less satised with
the choice process when, after learning about a product,
they realize that one of the options is actually not available
(as when a size in the middle of the range is not available
and cannot be tried on). They emphasize that such negative perceptions affect the stores image and might deter
future visits. Even more to the point, the empirical study
by Kalyanam et al. (2005) explores the implications of having key items within a product category, and conrm that
they deserve special attention. Their work also suggests that
stockouts of key items have a higher impact in the case
of apparel products compared to grocery stores. We also
observe that the inventory removal policy described above
guarantees that when a given article is being displayed in a
store a minimum quantity of it is exposed, which is desirable for adequate presentation. In that sense, the existing
studies on the broken assortment effect are also relevant
(see Smith and Achabal 1998 and references therein).2
Finally, we point out that our goal was to develop an
operational system for computing actual store shipment
quantities for a global retailer, as opposed to deriving
insights from a stylized model. Consequently, our model
formulation sacrices analytical tractability for realism, and
our theoretical contribution is small relative to that of the

seminal papers by Eppen and Schrage (1981) or Federgruen
and Zipkin (1984), for example. In fact, the key approximation that our optimization model formulation implements was derived in essence by Federgruen and Zipkin
(1984), whose analysis suggests that such approximation
leads to good distribution heuristics (see 3.2). On the other
hand, the present paper is the only one we are aware of
that presents a controlled pilot implementation study for
an inventory allocation model accounting for operational
details in a large distribution network (see 4). We also
believe that the simple performance evaluation framework
we developed when designing that study may be novel and
potentially useful to practitioners.
3. Model Development
In this section, we successively describe the two hierarchical models that we formulated to develop the optimization
software supporting the new process for inventory distribution discussed in 1. The rst (3.1) is descriptive and
focuses on the modeling of the relationship between the
inventory of a specic article available at the beginning of
a replenishment period in a single store and the resulting
sales during that period. The second model (3.2) is an
optimization formulation that embeds a linear approximation of the rst model applied to all the stores in the network to compute a globally optimal allocation of inventory
between them.
3.1. Single-Store Inventory-to-Sales Model
3.1.1. Store Inventory Display Policy at Zara. In many
clothing retail stores, an important source of negative customer experience stems from customers who have identied
(perhaps after spending much time searching a crowded
store) a specic article they would like to buy, but then cannot nd their size on the shelf/rack (Zhang and Fitzsimons
1999). These customers are more likely to solicit sales
associates and ask them to go search the back-room inventory for the missing size (increasing labor requirements),
leave the store in frustration (impacting brand perception),
or both. Proper management of size inventory seems even
more critical to a fast-fashion retailer such as Zara that
offers a large number of articles produced in small series
throughout the season. The presence of many articles with
missing sizes would thus be particularly detrimental to the
customers store experience.
We learned through store visits and personal communications that Zara store managers tend to address this challenge by differentiating between major sizes (e.g., S, M, L)
and minor sizes (e.g., XXS, XXL) when managing in-store
inventory. Specically, upon realizing that the store has run
out of one of the major sizes for a specic article, store
associates move all of the remaining inventory of that article from the display shelf/rack to the back room and replace
261
Figure 2.
Histogram of the adherence ratios DPFj /

DPAj across all stores.
50%
Fraction of Stores
it with a new article, thus effectively removing the incomplete article from customers sight.3 In contrast, no such
action is taken when the store runs out of one of the minor
sizes. The incomplete article that was removed might later
return to the oor if missing sizes can be shipped again
from the warehouse. Otherwise, it is either transferred to
another store where the sizes are consolidated, or remains
in the back room until clearance sales.
Zara does not have a product catalogue, and in fact
strives to maintain among its customers a sense of scarcity
and continuous assortment freshness (see 2). Consequently, customers do not typically enter a Zara store looking for a specic article, and do not expect articles not
displayed on shelves/racks to still be available in the back
room. The store inventory removal policy just described
can thus be seen as a balancing act between keeping inventory displayed to generate sales and mitigating the negative
impact of missing sizes on brand perception.
Interestingly, the denition of major and minor sizes may
reect that some sizes (e.g., M) account for considerably
more demand than others (e.g., XXL), but also more subtle psychological effects: When sizes XS, M, and L of a
given article are available but size S is not, for example,
S customers will tend to attribute that stockout to Zaras
mismanagement of its inventory. However, it appears that
size XS customers will place less blame on Zara when a
continuous set of sizes S, M, and L is available but XS is
not. This is because customers may not realize then that
some units of that article were made in size XS in the rst
place (not all articles are offered in extreme sizes), and
also because these customers may be blaming themselves
instead for their own seemingly unusual dimensions. As a
result, Zara managers seem to dene as major sizes either
a single size (e.g., M) or a continuous set of sizes (e.g., S,
M, L) in the middle of the size range, even in (common)
cases where an extreme size such as XS or XL accounts
for more demand than S or M.
We also learned that the inventory removal rule just
described was not prescribed by any formal policy imposed
upon store managers; rather, it constituted empirical observation of common store behavior. Because this seemed a
key modeling issue, we decided to verify through analysis how prevalent this policy was. Specically, in absence
of available data for whether in-store inventory is located
in the display area or in the back room, we measured the
adherence to the inventory display policy using the ratio
DPFj /DPAj for each store j, where DPAj is the number of
days, summed over all articles, when there was a stockout
of a major size but there was still some inventory available in another size, and DPFj corresponds to the subset
of those days characterized by the additional requirement
that no sales were observed for any size.4 The details of
this analysis are given in B of the online appendix, and
results are summarized in Figure 2, which shows the distribution of those ratios found across Zaras entire network
of approximately 900 stores (at the time when the data was
40%
30%
20%
10%
0%
0
0.2
0.4
0.6
0.8
DPF/DPA
collected). According to our metric, less than 2% of the

stores showed an adherence lower than 80%, with an average and median across stores both equal to 89%. We nd
these results to be quite compelling. In particular, they justify that the inventory display policy based on major sizes
can be used as a representation of store execution behavior.
We next describe a stochastic model developed to answer
the following question: Given the dependency between
inventory and sales of different sizes introduced by the
store inventory management policy based on major sizes,
how many sales of each article should be expected between
successive replenishments when starting from a given initial prole of inventory across sizes? As part of this rst
modeling effort, we initially assume away the dependencies
between inventory levels of different articles. That assumption is clearly not tenable in all cases because there may
in practice be signicant demand substitution (e.g., garments available in different colors but otherwise identical) and demand complementarity (e.g., assorted vest and
trousers sold separately) across articles. In E of the online
appendix, however, we discuss how our model may be
extended to the case of products offered in multiple colors.
3.1.2. Model Description. Consider an article offered
in a set of sizes = + , where + denotes the
major sizes (e.g., S M L) and the minor sizes
(e.g., XS XL). Sale opportunities for each size s
are assumed to be independent across sizes and follow
a Poisson process with rate s and cumulative counting measure Ns t, t 0, where t denotes the time
elapsed since the last replenishment (i.e., Ns t is the random number of sale opportunities for size s that occurred
between 0 and t). Although this Poisson assumption is primarily made for tractability reasons (the derivation to follow may be considerably harder, or even impossible, with
other demand process assumptions), we also believe that it
is a priori realistic, even though we were not able to obtain
data enabling a validation study. In addition, as mentioned
above, although there may in practice be some demand
dependencies across sizes (e.g., a customer preferring size
262
XS may still go for size S if XS is not available, customers

may choose the wrong size), we ignore these effects here.
Letting qs represent the inventory level of size s
immediately after replenishment at time 0 (that is, the sum
of any leftover inventory unsold in the previous period and
the quantity contained in the new shipment), we can now
dene a virtual stockout time s qs for every size s as
s qs inft 0 Ns t = qs
In words, s qs is the time at which, starting from an initial inventory of qs units, the store would run out of size s,
assuming that all inventory of that size remains always
exposed to customers and that no subsequent replenishment
ever occurs (these provisions justify the adjective virtual).
The earliest time at which one of the major sizes runs out,
assuming no replenishment occurs, can then be expressed
naturally as
+ q min+ s qs
s
In the following we will omit the dependence on the variables q = qs s when no ambiguity arises, and use the
notation a b mina b.
As described in 3.1.1, all inventory is removed from
customer view as soon as one of the major sizes runs out
at any point between successive replenishments. Under that
policy, the (random) total number of sales in a replenishment period resulting from an initial prole q of inventory
across sizes can be expressed as
Gq

s +
Ns + T +

s
Ns + s T
associated with sending a unit of inventory to different

stores, depending on how much inventory is already present
in these stores. Finally, the expected sales function should
also capture the complementarity effects across sizes following from the display inventory removal policy described
in 3.1.1. Specically, the marginal returns associated with
shipping inventory of one size should be nondecreasing
with the inventory level of the other sizes (the major sizes
in particular) because all sales processes terminate as soon
as a major size runs out.5 The expected sales function gq
associated with our model indeed exhibits those desirable
qualitative features, as formally established by the following proposition (where the notation es denotes a vector with
all components equal to zero except the sth, which is equal
to one).
Proposition 1. The expected sales function g is nondecreasing and discretely concave in each variable, and
supermodular. That is, gq is nondecreasing in xs and its
marginal differences s gq gq + es gq are nonincreasing in qs and nondecreasing in qs for all q
and s, s with s = s .
We turn next to the approximation. The rst step is
to note that each compensated Poisson process Ns t
Ns t s t denes a martingale with Ns 0 = 0, and that
the random variables + T and + s T appearing
in (1) are bounded stopping times. Doobs optional sampling theorem thus applies, and
g q = + + T +

s
(1)
where T > 0 denotes the time between consecutive replenishments (this would be one week for Zara). Given our purposes, we are particularly interested in characterizing the
expected value of the random sales function just dened,
or g q Gq, where the subscript s s highlights the dependence of that function on the demand
rate parameters characterizing the cumulative sales process
Ns s (we will omit that subscript when it is obvious
from context, however). In the following, we establish some
of its properties and develop an approximation for g that
may be easily embedded in a mixed-integer program (MIP).
3.1.3. Model Analysis and Approximation.
Intuitively, the descriptive model just dened captures how sales
should increase when more inventory is available for display in a store. Our expected sales function g should thus
obviously be nondecreasing with the inventory vector q.
A slightly less straightforward requirement is that function g should also reect the decreasing marginal returns
associated with shipping more inventory to a store, which
follows from demand saturation. This feature is particularly important given our ultimate goal because it will
effectively dictate the relative values of marginal returns
s + s T
where +
s
(2)
s +
Next, it follows from Jensens inequality that for any

subset of sizes ,
T min s T
s
(3)
In turn, the minimum operand in (3) can be calculated as

s T =
qs
qs

k s T
1
Ns T k =

s k=1
k=1 s k
(4)
where a
b a 1! is the Gamma function and
a b 0 va1 ev dv is the lower incomplete Gamma
functionthis follows from the optional sampling theorem
and properties of Poisson processes (see, e.g., Johnson et al.
1993). As is clear from (4), the expected stopping time
s T can be expressed as a sum of qs decreasing positive terms, so that it is a discretely concave function of qs .
That function is thus equal to the lower envelope of its
(discrete) tangents at every point, or
s T = minai s qs i + bi s
i
(5)
263

with ak s k s T /s k, bi s i1
k=0 ak s
for i 1 and b0 s 0 (we dene by extension
a
s 0 and b
s T ). Note that the parameter
ak s is equal to the average interarrival time weighted
by the probability that the (k + 1)th unit of size s will
sell before the next replenishment. Our proposed approximation consists of only computing the minimum in Equation (5) over a (small) nite subset instead of the
entire set of nonnegative integers . That is, we approximate function s T by the lower envelope of only a
few of its discrete tangents, thus obtaining an upper bound
for its exact value. Although that approximation can conceptually be made arbitrarily close (by considering a very
large number of discrete tangents), in practice we have
used small sets s dened as

s i
bi s 0 0
3T 0
6T
MIP

(6)
which are straightforward to compute numerically. That is,

our approximation of s T as a function of qs with
values in 0 T consists of the lower envelope of six tangents chosen so that their intercepts cover the range of
fractions of T shown in (6). Notice that this range of values is slightly denser as it gets close to T , that is, when
market saturation effects start to appear. Finally, substituting the approximate expression thus obtained for s T
in (3) for the sets = S + and = S + s, then substituting in turn the resulting expressions in (2), yields the
following approximation g for our original expected sales
function g :

s
s
min
s + s
i s

Pj zj + K
s.t.
ai s qs i + bi s
(7)
Note that g can be expressed as a linear combination of

minimums of linear functions of q, and may thus be easily
embedded in a MIP formulation having q as its primary
decision variables (as we proceed to do in the following
section). In addition, each of the two successive approximation steps (3) and (6) results in an upper bound of the original value, so that g q is also an upper bound for g q.
Finally, it is easy to see that the approximating function
g q still exhibits the desirable qualitative properties of
the original function g q stated in Proposition 1.
3.2. Network Sales Optimization Model
As stated in 1, our main objective is to implement an
optimization model for distributing a limited amount of
warehouse inventory between all stores in our industrial
partners retail network over time, with the goal of maximizing total expected revenue. As shown in Figure 1(b),
the primary input data of that model includes a demand

Ws
xsj Ws
jJ
zj

xsj
(8)
jJ
s
s
(9)

sj yj +
sj vsj
s
s +
j J
(10)
yj ai sj Isj + xsj i + bi sj

j J s + i sj
(11)
vsj ai sj Isj + xsj i + bi sj

j J s i sj
(12)
j J s
(13)
vsj yj
zj yj 0
g q = + min+ ai s qs i + bi s

s
i s
max

jJ
0
8T 0
9T T
forecast for every store as well as the inventory available in stores and in the warehouse. Note that, in practice,
this problem has a dynamic component because the shipment decisions in any given week impact future warehouse
and store inventory, and therefore both the feasible set of
shipments and the achievable sales in subsequent weeks.
In addition, this problem may involve connections between
different articles because of possible substitution or complementarity effects at the store level. For computational
and other practical reasons such as data availability, however, we have implemented a mixed-integer programming
(MIP) formulation that considers each article independently
(we return to this issue in E of the online appendix), and
only captures dynamic effects in a heuristic manner:
j J
vsj 0
s j J
xsj
s j J
(14)
In the formulation just stated, the primary decision variables xsj represent the shipment quantities of each size
s to each store j J for the current replenishment
period, which are constrained to be integer. These variables
are subjected to the warehouse inventory constraint (9),
which insures that the total shipment of a given size across
all stores never exceeds the inventory Ws available in the
warehouse for that size. The secondary decision variables zj
correspond to the approximate expected sales across all
sizes in each store j for the current period under consideration. The relationship assumed between these expected
sales and the shipment decisions xj xsj s , existing
store inventory Ij Isj s , and demand forecast j
sj s for each size s in each store j follows the approximate inventory-to-sales function derived in 3.1.3. Specifically, constraints (10)(13) along with the maximization
objective (8) insure that in any optimal solution to (MIP)
the variable zj is equal to gj xj + Ij , where g is the
approximate expected sales function dened in (7). The
264
expression for g is given by the right-hand side of constraint (10) and makes use of two auxiliary variables,
namely, yj and vsj , that are equal to

min+ ai sj qs i + bi sj
s
i sj
and

a

min
j qs i + bi s j
i
s

+
s s
i s j
respectively, in any optimal solution. The latter follows

from constraints (11)(13), and again from the fact that this
is a maximization problem.
The objective (8) is the sum of the expected revenue
in the current period and a heuristic approximation of
expected future revenue in subsequent periods. Its rst term
thus features the unit selling price Pj of the article considered in each store j, which constitutes input data. Note
that (as is typical in the retail industry) the selling price
may vary across stores, but is identical for all sizes sold
in the same store. The second objective term provides a
monetary evaluation of the total inventory remaining in the
warehouse after the shipment decisions considered are executed, using an exogenous unit value K for that inventory.
That value K can thus be interpreted as the unit opportunity cost of shipping inventory to a particular store and is
meant as a control lever allowing the model user to affect
its output: A high value of the warehouse inventory value K
relative to the store selling prices Pj results in conservative shipments, possibly appropriate shortly after a product
introduction (when forecast uncertainty is high), or when
the returns and transshipment costs associated with excessive inventory sent to low-selling stores may be particularly high. In contrast, a low relative value of K results
in aggressive shipments, perhaps suitable when forecasts
are deemed more reliable, and/or toward the end of the
planned shelf life of an article.
Note that any value of K above the selling price Pj
of a store will effectively prevent any shipments to that
store in any solution to (MIP) because any unit of store
inventory has a sales probability that is no larger than one
under our inventory-to-sales model (10)(13). More generally, K serves as a cutoff related to the expected revenue
associated with every marginal unit of inventory shipped to
a store j, itself equal to the sales price Pj times the sales
probability of that unit. For example, a value of K = 0
8Pj
will prevent the incremental shipment to store j of any
additional unit with an estimated sales probability lower
than 80%.
The warehouse inventory value K appearing in (8) thus
effectively allows the recommended shipments to reect
some of the dynamic considerations discussed earlier, even
though the model is otherwise myopic. Note that we do not
provide here any systematic method for deciding what the
appropriate value of K should be, leaving the determination of that control to the users judgement and experience
with the model. In addition, it is clear that the second term

in (8) is only a very crude approximation of the expected
revenue-to-go function that would appear in the objective of any dynamic programming formulation of the same
problem because it does not reect the existing distribution of inventory in stores, does not account for a possible unbalance of the warehouse inventory across sizes, etc.
However, our warehouse inventory value approximation
does constitute a simple implementation of an idea consistently described as fruitful in the literature when applied
to comparable stochastic inventory distribution models. In
particular, Federgruen and Zipkin (1984) found that decoupling the overall inventory distribution problem into a withdrawal decision (how much inventory in total should be
shipped to the stores) and an allocation decision (how that
inventory should be assigned to individual stores), then
solving the allocation problem in a myopic manner, led to
a good approximation (see also Zipkin 2000, Chapter 8 and
Gallego et al. 2007). Even though the policies proposed by
these authors for the withdrawal problem are obviously
more explicit and elaborated than our proposed withdrawal
solution, as described above, our model formulation otherwise implements the approximation scheme just described
fairly closely. In the optimal solution to (MIP), the value

of the overall quantity shipped s jJ xsj is determined
by the choice of K, and the individual shipments xsj also
solve the myopic allocation problem obtained when the

total withdrawal amount s jJ xsj is constrained to be
equal to that value.
4. Pilot Implementation Study

We were fortunate to help a team at Zara to implement
the new inventory allocation process described in earlier
sections, and test it as part of a small-scale but real pilot
experiment conducted during the 20062007 FallWinter
season (see the online appendix and Correa 2007 for more
details on that implementation and the software developed
to support it). This pilot test had three objectives: (i) prove
the feasibility of the new process envisioned through an
actual implementation; (ii) collect feedback from real users
to rene model features and the interface of supporting software; and (iii) estimate the specic impact of the new process on some key operational performance metrics. In this
section, we focus on the latter. Specically, we describe our
methodology in 4.1, discussing in turn the experimental
setup (in 4.1.1), the operational performance metrics used
(in 4.1.2), and our impact assessment method (in 4.1.3).
Results are then discussed in 4.2, and the nancial impact
is summarized in 4.3.
4.1. Methodology
4.1.1. Experimental Design. Because Zara is comprised of three subbrands or sections (women, men, and
children) that are organizationally distinct, it was felt that
265
the pilot study would be best organized within a single

one of them. Fifteen articles of the womens section were
thus selected by a senior Zara manager as a test bed for
the new process, with the goal of constructing as representative a sample as possible. Because articles at Zara are
split between a basic group (standard garments produced
in large quantities and sold during the whole season) and
a fashion group (trendier items with short life cycles produced in small batches), the relative proportions of these
two groups in the sample selected were in particular representative of the entire article population.
Our experimental setup leveraged the fact that Zara currently has only two major warehouses worldwide, the rst
in Arteixo (next to A Corua, northwest of Spain), shipping directly to about 500 stores in Western Europe, and
the second in Zaragoza (about halfway between Barcelona
and Madrid), shipping directly to about the same number
of stores located in the rest of the world.6 Specically, the
new inventory allocation process was only implemented in
Arteixo at some point during the life cycle of the 15 articles mentioned above, whereas the legacy process was still
used to distribute all articles (including these) in Zaragoza.
As further discussed in 4.1.3, our estimation of impact
associated with the new process is based on a comparison between that sample group of articles and a carefully selected control group of paired articles, conducted
with data from stores assigned to Arteixo. Because both
the sample and the control groups of articles were distributed using the legacy process in Zaragoza, by leveraging in turn the data from stores assigned to that second
warehouse we were also able to validate our estimation methodology. That is, we could quantify the likely
experimental error associated with our estimation of the
impact specically attributable to the new process because
any nonnegligible impact estimate in Zaragoza was obviously only attributable to such error, as opposed to the
new process.
For each article in the sample group, the warehouse team
in Arteixo thus switched from the legacy to the new process
at some point after July 2006, possibly after that article had
already been offered in stores for one or several weeks, and
kept using the new process until the end of the pilot, which
was set in November 2006 (December is not a representative month due to Christmas and the end of the season).
An important feature of that implementation is that the recommended shipments computed by the optimization model
were only presented as a suggestion to the warehouse team,
which retained the ability to freely modify them. We were
initially concerned that because of this discretion, any positive results would not be easily attributed to the new process. However, it turned out that very few modications
of the model output were actually performed after the rst
couple of model runs, which proved our apprehension to be
unfounded. Unfortunately, we were not able to choose the
exact week when the model would be used for the rst time
for each article. For that reason, in the end we were only
able to collect more than three weeks of data associated

with the new process for 10 articles out of the original 15,
and thus decided to remove the other 5 from our analysis.
However, among the remaining 10 were 4 basic and 6 fashion articles, which corresponds to approximately the same
proportions as in the overall assortment.
4.1.2. Operational Performance Metrics.
We now
present the framework we developed to measure the operational performance of Zaras inventory distribution over
time, and applied in particular to evaluate the impact
of our proposed process change. That framework is in
essence the same one that underlies the classical newsvendor problem, in that it captures the goal of neither shipping
too much nor too little inventory with respect to actual
demand. Specically, the two primary metrics we used,
to be described shortly, respectively measure any overstock (i.e., any amount of excess inventory shipped to the
stores) and understock (i.e., missed sales). In contrast with
the newsvendor model, however, these metrics have been
designed to assess the distribution of a large number of
articles across a network of many selling locations.
d
),
The primary data available to us included sales (Vrsj
d
shipments to store (Xrsj ), and returns to warehouse/
transshipments from store (Rdrsj ), all expressed in number
of units, on each day d of the entire study period, for each
available size s of each article r in a group of 118 (including
the pilot articles described in 4.1.1), for every Zara store
j in the world. Using basic inventory balance equations,
d
)
we derived the corresponding daily inventory positions (Irsj
w
and computed the corresponding weekly sales Salesrsj , shipments Shipmentswrsj , and returns/transshipments Returnswrsj
series by summing up the daily data series over each day d
in each (calendar) week w.7 Finally, we computed the corresponding networkwide cumulative weekly series Salestr ,
Shipmentstr , and Returnstr for each article r by summing up
the previous series over all stores j in the network, sizes s
of article r, and weeks w in the selling season up to and
including week t.
To quantify missed sales, we constructed data series
Demand wrsj and Demand tr , dened over the same index set
and providing estimates of uncensored customer demand,
that is, the sales that would have been observed had all merchandise been displayed without any stockout. Specically,
we rst computed
DNDwrsj

dw
d =0 or min
1Irsj
I d =0
sr+ r sj
and maxsr Vrdsj =0
or number of days in week w when size s of article r was

not on display at store j, either because that size was out
of stock, or because the article was removed due to the
inventory display policy described in 3.1.1. Second, we
estimated Demand wrsj by increasing sales according to the
number of days 7 DNDwrsj during which the item was
266
if Saleswrsj > 0 and DNDwrsj < 7 then

7
Demand wrsj = Saleswrsj
7 DNDwrsj
else
Demand wrsj = most recent nonnegative demand,
otherwise zero.
end
Note that our estimation of demand implicitly assumes
that average daily sales are identical throughout the week,
whereas many Zara stores do experience some predictable
variability within each week (e.g., surge of customer visits
on Saturday). We also ignore the time of the day in which
a stockout occurs, i.e, if an item sells out by noon, these
(reduced) sales still count as a whole day. Although the
resulting demand estimate could thus be biased, we do not
believe that this bias is likely to affect the new inventory
process and the old one in different ways (shipments occur
on a weekly basis according to a constant schedule), and
therefore feel that this simple approach is appropriate given
our purposes.
We used the ratio of cumulative sales to cumulative
shipments S/Srt Salestr /Shipmentstr , or shipment success
ratio, as our primary metric for quantifying any excess
inventory (i.e., overstock) in Zaras network. This represents the fraction of all units of a given article shipped to
stores since the beginning of the season that have actually been sold to date.8 At Zara, the shipment success ratio
S/Srt was actually used and closely monitored to evaluate
operational performance long before our collaboration. For
that reason, and also due to its convenience to estimate the
impact on sales (see 4.1.3), we chose this metric to measure overstock instead of inventory turns (or its reciprocal,
the average ow time), which might be more natural for
other retailers.
The primary metric we used for quantifying missed
sales due to lack of inventory (i.e., understock) is the
ratio of cumulative sales to cumulative demand S/Drt
Salestr /Demand tr , or demand cover ratio, where the
cumulative weekly demand series Demand tr is calculated
analogously to Salestr and Shipmentstr . This metric is to be
interpreted as the proportion of demand that Zara was able
to convert into sales through its display of inventory. In contrast with the rst metric, however, that second one was
new to Zara. We argued when introducing it that both were
required to form a comprehensive framework for evaluating
distribution performance, as illustrated by Figure 3. Observe
that whereas overstock and understock may not occur simultaneously in any inventory model describing the sales of a
given product in a single location (as the newsvendor), in
the network setting considered here both demand cover and
shipment success ratios may be low at the same time, as
explained in the lower-left quadrant of that gure.
Figure 3.
Shipment success ratio S/S

(cumulative Sales/cumulative Shipments)
actually on display and when those sales were observed,

according to the following procedure:
Proposed evaluation framework for Zaras

distribution performance.
High
Not enough
inventory
everywhere
Ideal
situation
Low
Too much
inventory in lowselling stores,
not enough in
high-selling stores
Too much
inventory
everywhere
Low
High
Demand cover ratio S/D

(cumulative Sales/cumulative Demand)
Besides the two primary metrics just discussed, we used

three additional secondary metrics. The rst is the Stock
Retention ratio, dened as SRtr 1 Returnstr /Shipmentstr .
This represents the fraction of units shipped to date that
were not sent to another store or sent back to the warehouse by the store manager who received them originally, and therefore provides an alternative measure for
overstock (although one that depends on the store managers actions).9 Our last two metrics are the Store Cover
ratio (SC tr ) and the Display Cover ratio (DC tr ), formally
dened as

d =0
wt
sr
jJ
dw 1Irsj
t
and
SC r 1
7 t r J

w
wt
sr
jJ DNDrsj
t
DC r 1
7 t r J
The store cover ratio is the fraction of cumulative days
sizes stores with stock at the store (possibly in the back
room), whereas the display cover ratio is the (smaller) fraction of these same days sizes stores with stock at the
store and in sufcient quantity to be displayed to customers, according to the store inventory policy described
in 3.1.1. They therefore both provide alternative measures for understock, although they are arguably coarser
than the demand cover ratio. This is because SC tr and
DC tr are inversely related to the number of days without
stock, as opposed to their economic consequence (i.e., the
number of units that could have been sold during those
days). These last two metrics therefore do not differentiate
between stockouts for the same period of time in high- and
low-selling stores, in contrast with the demand cover ratio,
but they give a sense of the service level.
Finally, note that values closer to one are more desirable for all the performance metrics just dened. In addition, the shipment success ratio S/S of a given article
tends to improve over time in Zaras environment, as
267
weekly sales progressively deplete the inventory already

shipped to stores and new shipments abate due to increasing warehouse inventory scarcity.10 The increasing inventory scarcity over a typical life cycle of a Zara article also
explains the natural tendency for the demand cover ratio
S/D (along with the other metrics SR, SC, and DC) to
decrease over time, as stockouts progressively occur earlier in the week following each replenishment and become
more widespread.
Several senior distribution managers at Zara interviewed
independently emphasized to us another important aspect
of the metric S/S, which is that, in their experience,
improving the shipment success ratio of a given article
becomes increasingly difcult for higher initial values of
that metric.11 Although we did not conduct a specic study
to investigate such an observation, we notice that it is plausibly explained by the broken assortment effect described
in Smith and Achabal (1998) and others (c.f. 2). Specically, a higher initial value of the shipment success ratio is
typically associated with lower inventory levels in stores,
so that the broken assortment effect could make it harder to
maintain or increase the S/S metric compared to a situation
with higher levels of available inventory.
The prevalence within Zara of that nonlinear notion of
performance relative to the metric S/S seemed potentially
problematic given our purposes. This is because our impact
assessment method (to be described in the next section)
involved the comparison of differences in the value of the
metrics dened over a given time period across distinct
articles. As a result, we decided to also consider as an alternative metric the logarithmic transformation ln1 S/S,
which seemingly captured the nonlinear behavior more
faithfully.12 Likewise, limiting the decrease of the demand
cover ratio and all other metrics from a given point in time
was perceived by Zara managers to be considerably more
challenging when the starting value of these ratios was
closer to one. As a result, we also considered the transformations lnM, where M is any of the other metrics S/D,
SR, SC, and DC (this alternative transformation was chosen
because these metrics tend to decrease over time).
4.1.3. Impact Assessment Method. Estimating the
impact of the new inventory process tested during the pilot
experiment on the metrics dened in 4.1.2 presents a
signicant but classical methodological challenge. Specifically, the most relevant basis for comparison, that is, the
values that these metrics would have taken for the pilot
articles over the same period of time had the new inventory process not been employed (i.e., the counterfactual),
cannot be directly observed. Our solution is known as
the difference-in-differences method, and is also used in
many other empirical event studies found in the literature
(e.g., Barber and Lyon 1996, Hendricks and Singhal 2005).
It involves using instead a control group as a basis for comparison, where that group is designed by carefully matching individuals in the population receiving the treatment
considered (in our case the articles included in the pilot

experiment) with others in the population at large (the articles still distributed using the legacy process).
For each one of the 10 pilot articles, we thus identify
a control article among the 118 articles included in our
data set that were distributed using the legacy process over
the same period of time. Our matching procedure is the
following: (1) a basic (respectively fashion) pilot article
may only be matched with a basic (respectively fashion)
control article (see 4.1.1); (2) dates when a pilot article
and its matched control were introduced cannot differ by
more than one week; and (3) subject to those restrictions,
the matched control article minimizes the initial difference
in performance with the pilot article, as measured by the
aggregate relative difference across shipment success and
demand cover ratios
S/Drt S/Drt
S/Srt S/Srt
+

t
t
maxS/Sr S/Sr maxS/Drt S/Drt
(15)
where r is the pilot article considered, r its matched control

article, and t is the week before the new process was used
for the rst time to distribute article r. That is, the notion of
proximity across articles that we use is driven by the values
of our primary performance metrics immediately before the
treatment begins (Barber and Lyon 1996 nd that matching
on such criteria leads to well-specied test statistics).
We carried out this matching procedure independently in
the Arteixo and Zaragoza warehouses for the 10 pilot articles (although the new process was only used in Arteixo;
see 4.1.1). For one article in Arteixo the control article initially selected was a clear outlier with an unusually
bad performance in the second half of the season (this
was conrmed by a standard box plot and Grubbs test),
resulting in an overly optimistic assessment of the new
process impact. We thus discarded that control article and
repeated the matching procedure. The outcome is summarized in Table 2.
Note that both performance ratios show signicant correlation among pilot and control articles in Arteixo. For that
warehouse the mean and median of the S/S ratios across
articles in the pilot and control groups are not statistically
different (p value > 0
1), but there is statistical evidence
showing that the S/D ratios are larger for the pilot articles.
In the case of Zaragoza, only the S/S ratios are signicantly correlated. The means and medians of the S/D ratio
and medians of the S/S ratios are not statistically different across pilot and control groups. Although the mean of
the S/S ratio is somewhat larger for the pilot articles, this
is not quite signicant (p value 0
1). Because the S/D
ratios are uncorrelated, we also performed the unpaired
tests and found that the mean and median were still not
different. Although such matching can never be perfect, we
believe ours to be suitable for our purposes, and we are in
particular unaware of any systematic bias that could make
the nal results be unduly optimistic. On the contrary, the
fact that the initial values of the S/D ratios are larger in
268
Table 2.
Outcome of the matching process at Arteixo and Zaragoza.
Number of pilot articles matched

Mean (median) shipment success ratio S/S of pilot articles (%)
Mean (median) shipment success ratio S/S of control articles (%)
Pearson (Spearman nonparametric) correlation coefcient
t-statistic (Wilcoxon signed-rank W -statistic) on the paired differences
Mean (median) demand cover ratio S/D of pilot articles (%)
Mean (median) demand cover ratio S/D of control articles (%)
Pearson (Spearman nonparametric) correlation coefcient
t-statistic (Wilcoxon signed-rank W -statistic) on the paired differences
Arteixo
Zaragoza
10
52
3
(46.8)
51
8
(52.1)
0
94 (0.89 )
0
19 (9)
62
0
(63.4)
53
4
(58.1)
0
85 (0.84 )
2
13 (39 )
10
50
0
(46.7)
46
5
(39.6)
0
98 (0.96 )
2
00 (33)
61
1
(55.4)
55
7
(53.8)
0
24 (0.08)
0
95 (21)
Note. The p values are two-tailed, except for the correlation coefcient, and the level of statistical signicance from zero is noted by p 01,
p 005, p 0025, p 001.
Arteixo is actually disadvantageous for the new process relative to the legacy process because this leaves less room
for improvement to the pilot articles. During our collaboration with Zara we were able to conrm that making
additional improvements to the ratios dened in 4.1.2 (or
equivalently, impeding their deterioration) is much more
challenging when the ratios are closer to one, as discussed
in 4.1.2.
Our next step is to compute the difference-in-differences
for each metric M dened in 4.1.2 and each matched pair
of articles r r
in each warehouse as
M MrEnd MrStart MrEnd MrStart
(16)
where MrStart (respectively MrStart ) is the value of the metric considered for the pilot (respectively control) article
the week before the new process was used for the rst
time, and MrEnd (respectively MrEnd ) is the corresponding
value at the end of the pilot experiment in November 2006.
For data relative to the Arteixo warehouse, the expression
in (16) thus provides an estimate for the specic impact
of the new process on the metric considered: The differences within parentheses excludes any time period other
than that when the new process was used, whereas the difference between the two pairs of parentheses is meant to
exclude the effects of factors other than the new process
(such as seasonality, weather, exceptional events), based
on the rationale that these external factors similarly affect
pilot and control articles. Because the legacy process was
used for both pilot and control articles in Zaragoza, expression (16), when calculated with data relative to that warehouse, provides an estimate of the error associated with
our impact assessment methodology (see 4.1.1). Expression (16), when computed for the S/S and S/D metrics,
can also be interpreted as a control-adjusted estimation of
the increase in sales attributable to the new process, relative
to either shipments (S/S) or demand (S/D). Rearranging
the terms dening S/S, for example, yields
S/S

SalesrEnd SalesrStart /ShipmentsrStart ShipmentsrEnd
=
ShipmentsrEnd
Start
/ShipmentsrStart
ShipmentsrEnd
SalesrEnd
SalesQ
ShipmentsrEnd

(17)
whereas the two terms in parentheses in (17), respectively,
correspond to the pilot article and the control article as
before, the numerator of each term represents the difference
between the actual nal cumulative sales and a proportional
prediction of what these sales would have been with the
legacy process, based on conditions immediately preceding the implementation of the new one. Because Salestr
minShipmentstr Demand tr , note that S/S and S/D
can thus also be interpreted as a somewhat conservative
estimate of the relative impact of the new process on sales.
4.2. Results
The results of the live pilot test are summarized in Tables 3
and 4. Our observations are based on averages across
articles of the values obtained for each metric using
Equation (16), which as discussed in 4.1.3 provides a
control-adjusted estimation for the impact of the new process on that metric. Note that considering averages across
articles is justied by the need to factor out the randomness (noise) that we cannot control (indeed, the focus on
such a statistic is prevalent in studies involving a pairwise matching procedure to construct a control group,
e.g., Hendricks and Singhal 2005). In addition, we report
associated t-statistics indicating whether these means are
signicantly different from zero, as well as the corresponding median for each metric and the respective nonparametric Wilcoxon signed-rank W -statistic (which likewise
indicates whether the median is signicantly different from
zero). The signicance of the statistics is reported conservatively by considering the two-tailed versions of the tests.
Notice that because our sample size is very small (only 10
articles), a difference from zero has to be fairly large for it
to be statistically signicant.
The results shown in Table 3 should be qualied in light
of the relatively low statistical signicance of the impact
on the primary metrics dened in 4.1.2, which is largely
driven by the limitation of sample size imposed upon us.
Nevertheless, Table 3 still suggests a positive impact on the
269
Table 3.
Final results of the live pilot test for the shipment success (S/S) and demand cover (S/D) ratios.
Original metrics
Arteixo
Mean (median) impact on basic articles (%)
Mean (median) impact on fashion articles (%)
Mean (median) impact on all articles (%)
t-statistic (W -statistic) on the models impact
Zaragoza
Mean (median) error for basic articles (%)
Mean (median) error for fashion articles (%)
Mean (median) error for all articles (%)
t-statistic (W -statistic) on the error
S/S
S/D
2
2 (1.8)
6
4 (7.4)
3
0 (0.6)
1
07 (17)
10
1 (8.6)
1
9 (2.0)
5
2 (7.9)
1
82 (35 )
5
3 (5.0)
0
5 (0.3)
2
4 (1.3)
0
76 (13)
2
6 (1.9)
4
6 (5.0)
3
8 (2.8)
1
55 (25)
Logarithmic transforms
ln1 S/S
15
0 (12.7)
18
6 (23.4)
17
1 (13.3)
3
17 (43 )
23
3
5
2
12
5
0
77
(21.5)
(0.2)
(0.2)
(11)
lnS/D
19
1
(18.8)
15
5
(9.4)
16
9
(17.9)
3
31 (47 )
8
6
8
4
8
5
1
54
(8.0)
(17.1)
(12.0)
(27)
p 005, p 0025, p 001.
primary metrics. These results are not driven by outliers

because the mean and median changes all have the same
sign. Considering all pilot articles, the changes in the value
of the S/S and S/D ratios in Arteixo are 3.0% and 5.2%,
respectively, whereas the corresponding estimation errors
given by measuring the same metrics in Zaragoza are
2
4% and 3
8%.13 The impact measured by the logarithmic transforms of these two metrics is even larger and different from zero with a high level of statistical signicance,
whereas the corresponding estimation errors obtained from
the Zaragoza warehouse are not. This latter set of results is
particularly noteworthy because the logarithmic transforms
of the S/S and S/D ratios constitute our most accurate
representation of Zaras managerial notion of performance.
Several interesting observations can be made by comparing the impact on each type of article in Arteixo (i.e.,
basic or fashion) with the respective estimation errors measured in Zaragoza. For basic articles, the mean impact on
the S/D ratio is positive (10.1%) and larger than the corresponding estimation error (2
6%), whereas the mean impact
on the S/S ratio is negative (2
2%) and smaller (in absolute terms) than its associated error (5.3%). In the case of
fashion articles, the mean impact on the S/S ratio is positive (6
4%) and larger than the corresponding estimation
error (0
5%), whereas the mean impact on the S/D ratio,
though still positive (1.9%), is smaller than its associated
error (4
6%). These results suggest that the new allocation
process impacts the two main types of articles in different ways. For basic articles its benets would mostly stem
from improvements in the demand cover S/D, whereas for
fashion articles they would consist of improvements in the
shipment success ratio S/S. These differences are plausibly
explained by the forecasting model developed as part of this
project (see 1), which, even though it is unbiased overall,
tends to slightly underestimate the demand of fashion articles and overestimate that of basic articles.14 Indeed, with
the new inventory distribution process, forecast underestimation errors conceptually generate insufcient shipments
favoring the S/S ratio to the detriment of the S/D ratio,
whereas overestimation errors generate excessive shipments
resulting in the opposite effect. Also supporting that interpretation is the observation that the correlation between the
individual S/S and S/D ratios of each article is negative
for both warehouses, but signicantly more so in Arteixo
(0
75), where the new inventory distribution process was
tested than in Zaragoza (0
40), which only used the manual legacy process. Unfortunately, we were not able to further investigate this issue because the forecasts used during
the pilot were not saved, and our attempt to reconstruct
them a posteriori were unsuccessful (the orders placed by
the store managers were not saved either).
Other reasons besides forecast biases may also explain
the different impact of the model on the primary metrics for
basic and fashion articles. In the case of the S/S ratio, the
apparently poor performance of the model for basic articles
has at least two alternative explanations: (i) the same two
(out of four) basic pilot articles that had negative S/S performance in Arteixo also performed badly (in fact, worse)
in Zaragoza, indicating that the choice of the basic pilot
articles was particularly adverse; and (ii) the initial values of the S/S ratio for the basic articles in the pilot was
relatively high (79
0% and 76
7% on average in Arteixo
and Zaragoza, respectively), making it harder for the model
to introduce signicant improvements (see related discussion in 4.1.2). Consistent with the latter explanation, note
that for basic articles in Arteixo the changes in mean and
median of the log transform of the S/S ratio relative to the
control articles (a metric designed to better reect performance independently of the starting point of the pilot) are
positive and signicantly different from zero, whereas the
corresponding estimation errors in Zaragoza are negative,
and not signicantly different from zero. In the case of the
demand cover ratio S/D for fashion articles, the outcome
of the matching process discussed in the previous section
may provide an alternative explanation for why the impact
measured in Arteixo is smaller than the estimation error.
In Arteixo, the initial value of the S/D ratio is larger for the
fashion pilot articles compared to the respective controls,
whereas in Zaragoza it is contrariwise. As in the previous
case, this seemingly negative result disappears when the
270
Figure 4.
Estimated relative change in sales for each pilot article in Arteixo and Zaragoza.
logarithmic transform of that ratio, which we deem to be

more meaningful, is considered instead.
We believe, however, that our most signicant results
stem from considering for each article the average of the
control-adjusted impact on the S/S and S/D ratios, i.e.,
S/S + S/D/2. As noted in the discussion of Equation (17), the control-adjusted impact estimations S/S
and S/D each provide an estimation of the relative
change in sales attributable to the model for each article
in the pilot test, based on the assumption that the S/S
and S/D ratios would have remained relatively unchanged
over the test period under the counterfactual scenario.
However, the strong negative correlation between S/S
and S/D for each article, which is plausibly caused
by forecast biases as discussed above, suggests that their
average provides an estimate of the same quantity, which
is arguably more robust. Figure 4, which contains a plot
of these averages for each pilot article distributed from
Arteixo and Zaragoza, is striking in that respect: According
to that measurement, the relative sales impact in Arteixo
is positive for every single article, with a mean across
articles of 4
1% (median 4
2%), whereas the corresponding estimation error calculated using data from Zaragoza
is centered around zero (mean and median across articles are 0
7% and 0
6%, respectively).15 Subtracting
this estimated experimental error suggests that the relative impact of the new distribution process on sales is of
the order of 3%4%, which is substantial for a retailer
like Zara, given that this improvement does not appear to
require any major costs.
Table 4.
The results reported in Table 4 are similar to those discussed for the primary metrics. The impact on the stock
retention ratio SR in Arteixo is larger than the estimation
error, in particular for basic articles, suggesting that the
model effectively reduces the level of transshipments. The
measured impact of the new process on the store cover SC
and display cover DC ratios is also positive overall. However, this result is only driven by basic articles because
for fashion articles the impact on these ratios (and their
log transforms) remains just under the estimation error. We
note that basic articles have life cycles that sometimes span
the whole season, whereas fashion articles are by design
only sold in stores for a few weeks, which may be why
improving their store cover and display cover ratios seems
more difcult. The fact that the overall impact is greater
on the DC ratio than on the SC ratio is noteworthy, however, because a distinguishing feature of our model is precisely to capture the display of inventory on store shelves
and racks (display cover), as opposed to its presence anywhere in the store, including the back room where it does
not sell (store cover). As before, the results are not driven
by outliers because the mean and median changes are
thoroughly consistent; the results in Arteixo are signicantly different from zero, whereas the estimation errors are
not; and the statistical signicance of the impact improves
when the logarithmic transforms of the performance metrics are considered instead.
An interesting question is whether the results just
described can be attributed to any consistent qualitative
Final results of the live pilot test for the store retention (SR), store cover (SC), and display cover (DC) ratios.
Original metrics
SR
SC
Arteixo
Mean (median) impact on basic articles (%)
0
8 (0.8)
6
1 (6.4)
Mean (median) impact on fashion articles (%) 0
7 (0.0)
1
1 (0.6)
Mean (median) impact on all articles (%)
0
7 (0.2)
3
1 (3.7)
t-statistic (W -statistic) on the models impact
2.17 (31 )
2
19 (37 )
Zaragoza
Mean (median) error for basic articles (%)
0
0 (0.0)
3
1 (2.7)
Mean (median) error for fashion articles (%) 0
5 (0.4) 1
7 (1.7)
Mean (median) error for all articles (%)
0
3 (0.2) 2
3 (1.8)
t-statistic (W -statistic) on the error
1
24 (21) 1
65 (27)
Logarithmic transforms
DC
lnSR
lnSC
lnDC
7
5 (7.6)
2
1 (1.8)
4
3 (5.0)
2
59 (41 )
0
8 (0.8)
0
7 (0.0)
0
8 (0.2)
2
19 (31 )
7
3 (8.1)
1
5 (0.8)
3
8 (4.1)
2
38 (39 )
9
4 (9.8)
3
2 (2.5)
5
7 (6.1)
2
82 (43 )
3
8
2
0
2
7
1
72
4
4
3
3
3
7
1
76
3
5
2
5
2
9
1
64
(3.2)
0
0
(2.9) 0
5
(2.9) 0
3
(29)
1
17
(0.0)
(0.4)
(0.2)
(21)
(3.3)
(2.0)
(2.1)
(27)
(4.0)
(3.6)
(3.6)
(31)
p 005, p 0025, p 001.
271
differences between the shipment decisions generated by

the new optimization-based distribution process and those
resulting from the manual legacy process. Although we
have not been able to collect sufcient data to perform a
systematic study of this issue, the following are the most
obvious insights that we obtained after discussing the solutions with the warehouse team.
A typical rst reaction by warehouse team members
when inspecting the shipments recommended by the model
was that these decisions seemed reasonable overall, and
that in many cases they would have entered a quantity that
differed by a couple of units at most.
There were also situations when the solution seemed
counterintuitive to them at rst. One frequent source of
discrepancy stemmed from the shipment decisions for different sizes of the same article. Given the data-intensive
decisions that the warehouse team had to make under time
pressure, in some cases they would not fully account for
the dependencies across sizes (see 3.1.1) and, in contrast
with the model, would occasionally recommend shipments
resulting in a size prole likely to be removed from display
relatively quickly (if not immediately).
We observed that when the remaining stock of an item
was low, the manual practice typically consisted of trying
to satisfy the requests of as many stores as possible. In contrast, the model tended then to avoid stock fragmentation
by shipping to a smaller number of stores where the inventory would sell with greater probability and where full sets
of sizes could be completed.
Another source of discrepancy occurred when the
shipment requests from a particular store seemed high in
light of its inventory level and recent sales (perhaps that
store anticipated some inventory rationing in subsequent
weeks). The warehouse team would then typically cut down
the shipment request, but did not have a formal rule to
decide by how much, and as a result tended to reduce the
requests by a signicantly lower amount than the model.
4.3. Financial Impact
We conclude this section with an assessment of the nancial impact of the new process.16 Following the pilot study,
Zara made the decision to use our new allocation process for all its articles and stores. This large-scale deployment effectively occurred over the rst half of 2007. On
September of the same year, Zaras CFO estimated that the
nancial benets specically attributable to the use of our
model were consistent with the sales increase reported here
for the live pilot.
If we take 2006 as the base year, and assuming that the
model increased revenues by 3.5% in 2007 (see 4.2), we
thus estimate that the new process generated about $275 M
in additional revenue or $34 M in additional net income
for 2007.17 18 Given the nding from the pilot experiment
that the model does not increase average inventory throughout the retail network (see 4.2), this additional net income
corresponds to an increase of the return on assets by at least
3.5% (note that the new process does not affect assets other
than inventory).
The nancial impact estimation just stated ignores
any changes in operational costs that might have been
introduced by the new process. This seems conservative
because the software license and labor costs incurred during the implementation (see below) appear smaller than the
reduction in warehouse returns and transshipments costs
measured during the pilot study (4.2), which the previous
estimate based on additional revenues also ignores. We note
as well that Zaras earnings before taxes and interest (EBIT)
amounted to 16.5% of revenues in 2006. Ignoring again any
impact on operational costs, a 3.5% revenue increase due to
the new process would imply a 21.3% increase in EBIT for
2007. The actual increase from 2006 to 2007 was 22.5%,
which at least to some extent validates our estimations.
Finally, the main cost associated with this implementation
stemmed from the development time spent by the project
team members (in comparison, the software costs represented a minor expense, see F in the online appendix).
Although some team members had several simultaneous
assignments, we estimate that the total time invested in the
project is equivalent to one manager experienced in distribution and store management and two engineers experienced
in databases and optimization software working together
full time during one year, which would represent less than
$1 M in labor costs. Given the nancial impact estimated
above, this would correspond to a return on investment for
the project after one year amounting to 30 times or more.
5. Conclusion
This paper involved the development of a new operational
process to allocate scarce inventory across the store network of a fast-fashion retailer. The most salient feature
of that process is arguably its reliance on an optimization model capturing inventory display policies at the store
level. In addition, we also reported the implementation and
test of that new process as part of a live pilot experiment,
using a performance evaluation framework that may be
of independent interest. The results of the live pilot test
suggest that the new process increases sales (by 3%4%
according to our best estimate), decreases transshipments,
and increases the proportion of time that the articles are
on display. As of the time of writing, every item currently
found in any Zara store worldwide has been shipped to
that store based on the output of the optimization model
described in 3.2 of this paper. In addition, the Inditex
group is also planning to start using that new process for
its other brands in the near future.
Beyond its nancial impact, this new allocation process
has also had organizational implications that we believe
are positive. In particular, the warehouse allocation team
has seen its responsibilities shift from repetitive data entry
towards exception handling, scenario analysis, process performance evaluation, and improvement. That team deserves
272
special recognition in our view, for it has played a pivotal
role in the improvement and successful implementation of
the new allocation model, and has demonstrated to us the
importance of human experience when facing many distribution issues. To the best of our knowledge, Zara is not
planning to leverage any productivity gains associated with
the new allocation process through head-count reductions;
however, we expect the new process to generate substantial
economies of scale if the company continues to grow as
planned. In addition, store managers may be asked in the
future to provide input to forecasts as opposed to shipments
(see Correa 2007).
This project may also have had some cultural impact
at Zara, a company that we believe owes part of its success to the unique intuition of its founder. We doubt that
Zara will ever use advanced mathematical models to help
with several of its key challenges, including anticipating
volatile market trends, recruiting top designers, and creating fashionable clothes, and it is not clear to us that it
should. In fact, we see the story of Zaras success as a humbling one given our background, because a key aspect of
its business model is to leverage the endogenous increase
in demand associated with short product life cycles, a
feature not predicted or quantied by any of the current
quantitative inventory-purchasing models that we know of
(Fisher et al. 2000). However, this collaborative interaction
may have inuenced Zaras realization that for other processes involving large amounts of quantitative data, such as
distribution and pricing, formal operations research models may lead to better performance and more scalable
operations.
Beyond Zara, we expect that our model may also be
useful to other retailers managing a network of stores, particularly those facing lost sales (as opposed to backorders)
and dependencies across sizes introduced by store display
policies. Indeed, the latter feature has not received much
attention in the literature, and the present work suggests
that accounting for it may have a signicant impact on
sales. In terms of future work, the methodology we applied
(solving a large-scale industrial optimization problem subject to uncertainty by embedding the linear approximation
to a stochastic performance evaluation model in an MIP
formulation) may be applicable to other contexts beyond
retailing. Further related theoretical work could thus be
interesting, for example, characterizing the suboptimality of
our approximate MIP formulation, or the development of a
unied framework for general allocation problems. Finally,
we see other research opportunities motivated by the specic features of fast-fashion retailers relative to traditional
retailers. In particular, further investigations of store-level
inventory display policies and warehouse ordering policies
seem warranted.
6. Electronic Companion
An electronic companion to this paper is available as part
of the online version that can be found at http://or.journal.
informs.org/.
Endnotes
1. See Chen (2005) for a study on compensation packages
that induce salespeople to forecast accurately and to work
hard.
2. The broken assortment effect refers to the empirical
observation that the sales rate for an article decreases as
the inventory available diminishes, even though that inventory may still be positive. This is explained by the reduced
visibility of that article to customers in the store then, and
the fact that popular sizes and colors of that article may
become progressively harder to nd.
3. Weekly shipments to a Zara store include 10%20% of
new articles. If a new article is not available to replace an
incomplete one, the latter is still removed, but the store
manager rearranges the articles remaining on the oor to
avoid large empty spaces.
4. DPA and DPF stand for Days when the Policy was
Applicable and Days when the Policy was Followed,
respectively.
5. This prediction is consistent with our assumption that
any substitution effects across sizes driven by customer
behavior are negligible relative to the complementary effects
driven by the inventory display policy described. Although
relaxing that assumption seems an interesting research challenge, we do not explore this further here.
6. Because stores in Western Europe tend to be more established and sell more merchandise, the Arteixo warehouse
currently ships roughly 75% more volume than the one in
Zaragoza.
7. Point-of-sale data tends to be very accurate at Zara,
but data inconsistencies concerning the inventory position
could be more frequent. Although we were not able to precisely estimate these inconsistencies during the pilot, we
measured impact based on a differential relative to control
articles (see 4.1.3 below) and there is no reason why any
data inaccuracy concerning the inventory position would
selectively affect the pilot articles more than the control
ones.
8. The value of the S/Srt metric at the end of the season
thus corresponds to the widely used metric of sell-thru or
retail turnover, which is dened as the fraction of sales
to total inventory received by suppliers.
9. At Zara, store transshipments and returns to the warehouse also require the approval of the regional manager.
10. From now on we omit the indices t and r when no
ambiguity arises.
11. For example, increasing the shipment success ratio of
an article from 0.8 to 0.9 is regarded within Zara as a considerably superior achievement by a distribution manager
than an increase of that ratio for another article from 0.4
to 0.5 over the same period of time.
12. Continuing the previous example, changes of S/S from
0.4 to 0.5 and from 0.8 to 0.9, respectively, correspond to
increases of 0.2 and 0.7 for ln1 S/S.
13. Average ow time is an alternative measure of overstock, and we observed that its control-adjusted value

decreased by almost one week at Arteixo, whereas at

Zaragoza it remained practically the same.
14. This likely stems from the different demand patterns
and amount of input data available for basic and fashion
articlessee Correa (2007) for more details on the forecasting model as well as its accuracy.
15. The difference between Arteixo and Zaragoza is statistically signicant with p values less than 0.05 and 0.065
for the mean and median, respectively.
16. The nancial impact estimations provided here were
performed independently by the paper authors and did not
engage the responsibility of the Inditex group.
17. See C of the online appendix for the extract
calculation.
18. The validity of extrapolating the specic sales increase
measured during the pilot study to subsequent periods
is subject to a number of assumptions, including the
relative stability of Zaras products attractiveness to its
customers.
Acknowledgments
The authors rst thank their industrial partner, Zara,
and their two key collaborators, Jos Antonio Ramos
Calamonte and Juan Correa. They are also particularly
grateful to other Zara employees, including Javier Garca,
Miguel Daz, and Jos Manuel Corredoira (Pepe). The
authors also thank the area editor, the associate editor, and
two anonymous referees for many comments that helped
improve the paper. They refer the reader to the online
appendix for acknowledgments of other contributing individuals and funding sources.
References
Axster, S., J. Marklund, E. A. Silver. 2002. Heuristic methods for centralized control of one-warehouse, N -retailer inventory systems. Manufacturing Service Oper. Management 4(1) 7597.
Barber, B. M., J. D. Lyon. 1996. Detecting abnormal operating performance: The empirical power and specication of test statistics.
J. Financial Econom. 41 359399.
Cachon, G., M. Lariviere. 1999. Capacity choice and allocation: Strategic behavior and supply chain performance. Management Sci. 45(8)
10911108.
Caro, F., J. Gallien. 2007. Dynamic assortment with demand learning for
seasonal consumer goods. Management Sci. 53(2) 276292.
273
Chen, F. 2005. Salesforce incentives, market information, and production/inventory planning. Management Sci. 51(1) 6075.
Correa, J. 2007. Optimization of a fast-response distribution network. M.S.
thesis, LFM, Massachusetts Institute of Technology, Cambridge, MA.
Eppen, G., L. Schrage. 1981. Centralized ordering policies in a
multi-warehouse system with leadtimes and random demand.
L. Schwarz, ed. Multi-Level Production/Inventory Control Systems:
Theory and Practice. North Holland, Amsterdam, 5169.
Federgruen, A. 1993. Centralized planning models for multi-echelon
inventory systems under uncertainty. S. C. Graves, A. H. G. Rinnooy Kan, P. H. Zipkin, eds. Logistics of Production and Inventory.
Handbooks in Operations Research and Management Science, Vol. 4.
North-Holland, Amsterdam, 133173.
Federgruen, A., P. Zipkin. 1984. Approximations of dynamic multilocation
production and inventory problems. Management Sci. 30 6984.
Fisher, M. L., A. Raman, A. S. McClelland. 2000. Rocket science retailing
is almost hereAre you ready? Harvard Bus. Rev. 78(4) 115124.
Fraiman, N., M. Singh, L. Arrington, C. Paris. 2002. Zara. Columbia
Business School Case, Columbia University, New York.
Gallego, G., . zer, P. Zipkin. 2007. Bounds, heuristics, and approximations for distribution systems. Oper. Res. 55(3) 503517.
Ghemawat, P., J. L. Nueno. 2003. ZARA: Fast fashion. Harvard Business School Multimedia Case 9-703-416, Harvard Business School,
Boston.
Graves, S. C. 1996. A multiechelon inventory model with xed replenishment intervals. Management Sci. 42(1) 118.
Hendricks, K. B., V. R. Singhal. 2005. Association between supply chain
glitches and operating performance. Management Sci. 51(5) 695711.
Jackson, P. L. 1988. Stock allocation in a two-echelon distribution system or what to do until your ship comes in. Management Sci. 34(7)
880895.
Johnson, N., S. Kotz, A. Kemp. 1993. Univariate Discrete Distributions.
Wiley Interscience, New York.
Kalyanam, K., S. Borle, P. Boatwright. 2005. Modeling key item
effects. Working paper, Tepper School of Business, Carnegie Mellon
University, Pittsburgh.
McGavin, E. J., L. B. Schwarz, J. E. Ward. 1993. Two-interval inventoryallocation policies in a one-warehouse, N -identical-retailer distribution system. Management Sci. 39(9) 10921107.
Muckstadt, J. A., R. O. Roundy. 1993. Analysis of multistage production systems. S. C. Graves, A. H. G. Rinnooy Kan, P. H. Zipkin,
eds. Logistics of Production and Inventory. Handbooks in Operations Research and Management Science, Vol. 4. North-Holland,
Amsterdam, 59131.
Smith, S. A., D. D. Achabal. 1998. Clearance pricing and inventory policies for retail chains. Management Sci. 44(3) 285300.
Zhang, S., G. J. Fitzsimons. 1999. Choice-process satisfaction: The inuence of attribute alignability and option limitation. Organ. Behav.
Human Dec. 77(3) 192214.
Zipkin, P. 2000. Foundations of Inventory Management. McGraw-Hill,
Burr Ridge, IL.

Inventory Management of A Fast-Fashion Retail Network: Felipe Caro Jérémie Gallien

Uploaded by

Copyright:

Available Formats

Inventory Management of A Fast-Fashion Retail Network: Felipe Caro Jérémie Gallien

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Inventory Management of A Fast-Fashion Retail Network: Felipe Caro Jérémie Gallien

Uploaded by

Copyright:

Available Formats

informs

Inventory Management of a Fast-Fashion

only 2,0004,000 items for key competitors. This increases

The recent impressive nancial performance of the Spanish

Caro and Gallien: Inventory Management of a Fast-Fashion Retail Network

Operations Research 58(2), pp. 257273, 2010 INFORMS

(b) New process envisioned

Past sales data

reference and size

Requested shipment quantities

that although the general problem just stated is not specic

and career promotion prospects are driven to a signicant

Caro and Gallien: Inventory Management of a Fast-Fashion Retail Network

Operations Research 58(2), pp. 257273, 2010 INFORMS

Caro and Gallien: Inventory Management of a Fast-Fashion Retail Network

Operations Research 58(2), pp. 257273, 2010 INFORMS

our theoretical contribution is small relative to that of the

Caro and Gallien: Inventory Management of a Fast-Fashion Retail Network

Operations Research 58(2), pp. 257273, 2010 INFORMS

Histogram of the adherence ratios DPFj /

collected). According to our metric, less than 2% of the

Caro and Gallien: Inventory Management of a Fast-Fashion Retail Network

Operations Research 58(2), pp. 257273, 2010 INFORMS

XS may still go for size S if XS is not available, customers

associated with sending a unit of inventory to different

Next, it follows from Jensens inequality that for any

In turn, the minimum operand in (3) can be calculated as

Caro and Gallien: Inventory Management of a Fast-Fashion Retail Network

Operations Research 58(2), pp. 257273, 2010 INFORMS

which are straightforward to compute numerically. That is,

ai s qs i + bi s 

Note that g can be expressed as a linear combination of

yj ai sj Isj + xsj i + bi sj 

vsj ai sj Isj + xsj i + bi sj 

g q =  + min+ ai s qs i + bi s 

Caro and Gallien: Inventory Management of a Fast-Fashion Retail Network

respectively, in any optimal solution. The latter follows

Operations Research 58(2), pp. 257273, 2010 INFORMS

with the model. In addition, it is clear that the second term

4. Pilot Implementation Study

Caro and Gallien: Inventory Management of a Fast-Fashion Retail Network

Operations Research 58(2), pp. 257273, 2010 INFORMS

the pilot study would be best organized within a single

able to collect more than three weeks of data associated

and maxsr Vrdsj =0

or number of days in week w when size s of article r was

Caro and Gallien: Inventory Management of a Fast-Fashion Retail Network

if Saleswrsj > 0 and DNDwrsj < 7 then 

Shipment success ratio S/S

actually on display and when those sales were observed,

Operations Research 58(2), pp. 257273, 2010 INFORMS

Proposed evaluation framework for Zaras

Demand cover ratio S/D

Besides the two primary metrics just discussed, we used

Caro and Gallien: Inventory Management of a Fast-Fashion Retail Network

Operations Research 58(2), pp. 257273, 2010 INFORMS

weekly sales progressively deplete the inventory already

considered (in our case the articles included in the pilot

ai s qs i + bi s

yj ai sj Isj + xsj i + bi sj

vsj ai sj Isj + xsj i + bi sj

g q = + min+ ai s qs i + bi s

if Saleswrsj > 0 and DNDwrsj < 7 then