informs ®

Vol. 40, No. 1, January–February 2010, pp. 71–84 doi 10.1287/inte.1090.0472

issn 0092-2102 ! eissn 1526-551X ! 10 ! 4001 ! 0071 © 2010 INFORMS

THE FRANZ EDELMAN AWARD

Achievement in Operations Research

Zara Uses Operations Research to Reengineer Its

Global Distribution Process
Felipe Caro
UCLA Anderson School of Management, University of California, Los Angeles, Los Angeles, California 90095,
fcaro@anderson.ucla.edu
Jérémie Gallien
MIT Sloan School of Management, Massachusetts Institute of Technology, Cambridge, Massachusetts 02142,
jgallien@mit.edu

Miguel Díaz, Javier García, José Manuel Corredoira, Marcos Montes

Zara, 15142 Arteixo, La Coruña, Spain {migueldm@inditex.com, javiergt@inditex.com, josemcc@inditex.com, marcosmv@inditex.com}

José Antonio Ramos

Carrefour, 28028 Madrid, Spain, joseantonio_ramos_calamonte@carrefour.com

Juan Correa
Dell Computers, Austin, Texas 78759, jcorrea@sloan.mit.edu

Overcoming significant technical and human difficulties, Zara recently deployed a new process that relies exten-
sively on sophisticated operations research models to determine each inventory shipment it sends from its two
central warehouses to its 1,500 stores worldwide. By taking a retail size-assortment view of a store’s inven-
tory, the model incorporates the link between stock levels and demand to select store replenishment quanti-
ties. Through a rigorous, controlled field experiment, we estimate that this new process has increased sales by
3–4 percent; this corresponds to estimated profits of approximately $233 million and $353 million in additional
revenues for 2007 and 2008, respectively.
Key words: retailing; fast fashion; inventory management; field experiment; sizes distribution.

Introduction offered in appealing store environments and at com-

petitive prices.
W ith more than 1,500 stores in 68 countries and
E6.26 billion in 2007 annual sales, Zara is one
To support this customer-value proposition, Zara
has developed an innovative and highly respon-
of the world’s leading fashion retailers and the flag-
sive design, production, and distribution infrastruc-
ship chain of the Inditex Group. Because of its impres-
ture that many press articles and case studies (e.g.,
sive growth in recent years, Zara has also become Fraiman et al. 2002, Ghemawat and Nueno 2003) have
one of the most recognized apparel brands worldwide described. In particular, Zara’s supply chain involves
(Helm 2008). This success is widely attributed to its two primary warehouses in Spain that periodically
fast-fashion business model, which involves frequent receive shipments of finished clothes from suppli-
in-season assortment changes and ever-trendy items ers and ship merchandise directly to each Zara store
71
 Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
72 Interfaces 40(1), pp. 71–84, © 2010 INFORMS

worldwide twice a week (at the time of this writ- Remaining

ing, Zara had opened two additional warehouses in sizes Action

Madrid and León, Spain). This paper discusses the S M L Keep on display
development, implementation, and impact of a new S M L Keep on display
OR-based process; since 2006, Zara has used it daily to
M L Keep on display
determine these shipment quantities, i.e., the “blood-
stream” to its only sales channel. It constitutes a S M Keep on display
less technical and more practice-oriented companion M Keep on display
paper to Caro and Gallien (2009).
S L Move to backroom
Our work began in mid-2005 from a contact that
Caro (University of California, Los Angeles) estab- S Move to backroom
lished with Ramos (then working at Zara). It soon L Move to backroom
evolved into a collaboration of researchers Caro and
Gallien (Massachusetts Institute of Technology, MIT)
Figure 1: For an article offered in three sizes with M (medium) as the only
with Zara employees Ramos, García, and Montes major size, any combination without size M is moved to the backroom.
under the executive supervision and sponsorship of
Zara’s CFO, Miguel Díaz. Inditex senior informa-
tion technology (IT) engineer Corredoira had over- To distribute merchandise to its stores, Zara uses
all responsibility for the system implementation; the a supply chain that consists of two primary ware-
project also leveraged the six-month internship of houses in Spain. They periodically receive shipments
(then) MIT graduate student Correa hosted by Indi- of finished clothes from suppliers and ship replenish-
tex, which initiated a corporate partnership with ment inventory directly to each Zara store worldwide
MIT’s Leaders for Manufacturing (LFM) Program twice a week. The key associated control challenge is
(Correa 2007). to determine the exact number of units of each size
In this paper, we discuss the managerial problem (up to eight) of each article (up to 3,000 at any time)
we addressed, the solution we developed, and the that should be included in each shipment to each of
impact of our work, and we end with a summary its more than 1,500 stores. This control problem is par-
of our contributions to OR practice and some con- ticularly challenging for the following reasons:
cluding remarks. We emphasize that the authors per- • As Figure 1 illustrates, most stores only display
formed all the financial impact estimates discussed in merchandise for sale when the set of available sizes
this paper. They did not engage the responsibility of is complete enough. Their intention is to achieve
the Inditex Group, which advises that any forward- a balance between keeping inventory displayed to
looking statement is subject to risk and uncertainty generate sales and mitigating the impact of missing
and could thus differ from actual results. sizes on brand perception; this is driven by the neg-
ative feeling that customers experience when they
The Problem have identified a specific article they would like to
buy, perhaps after spending much time searching a
Managerial Challenge
Zara’s innovative business model is powered by a crowded store only to learn that their size is not avail-
continuous cycle that involves flows from stores to able. More specifically, store managers tend to differ-
designers (market information and customer desires), entiate between major sizes (e.g., S, M, L) and minor
designers to suppliers (production orders for new sizes (e.g., XXS, XXL) when managing in-store inven-
designs), suppliers to warehouses (deliveries of fin- tory. When a store runs out of a major size for a
ished clothes), and warehouses to stores (outbound specific article, store associates move all remaining
merchandise shipments). This last link is particularly inventory of that article from the display area to the
critical; it constitutes the bloodstream of Zara’s mer- backroom and replace it with a new article, effectively
chandise to its unique sales channel and directly removing the incomplete article from customer sight.
affects Zara’s global revenues. In contrast, they take no such action when the store
Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
Interfaces 40(1), pp. 71–84, © 2010 INFORMS 73

runs out of a minor size. They might return a previ- • The amount of relevant data (warehouse inven-
ously removed article back to the floor if the missing tory, store inventory, and store sales history for each
sizes can be shipped again from the warehouse; other- article) is enormous.
wise, it is either transferred to another store where the • The available warehouse inventory is often
sizes are consolidated or it remains in the backroom limited. This follows from Zara’s business model,
until the store has a clearance sale. Note that this store whereby the store life cycle of articles typically spans
inventory display policy introduces significant ship- only a small fraction of a selling season (i.e., five to
ment interdependencies across sizes; it may be point- six weeks), so that the store assortment turns over
less to ship some units of a given size (e.g., XS) if it much more frequently than the assortment of more
is not accompanied by enough units of a major size traditional retailers does.
(e.g., M) to trigger display. It is worth noting that the
removal rule described above is not prescribed by any Legacy Process
formal policy imposed upon store managers; it con- Until 2006, Zara exclusively used the legacy pro-
stitutes an observation of common store behavior that cess, as Figure 2(a) shows, to generate all its store-
we validated empirically (Caro and Gallien 2009). replenishment shipment decisions worldwide. As
• These shipment decisions must be determined in part of that process, store managers received weekly
only a few hours after the relevant information (e.g., statements showing the subset of articles available in
current store inventory, previous-day sales history) the central warehouse for which they could request
becomes available. Any further delay, in light of ware- shipments to their stores. Note that these weekly
house processing times and transportation schedules, statements (dubbed “the offer”) would thus effec-
would effectively delay the replenishment of stores by tively implement any high-level assortment decisions
one full day (this replenishment response time is par- made by Zara’s headquarters for each store. How-
ticularly important to Zara’s business model, which ever, a statement would not mention the total quan-
explains why direct shipments to all stores are sent tity of inventory available in the warehouse for
by truck and air every week). each article listed. After considering the inventory
• The number of associated shipment decisions remaining in their respective stores, store managers
reaches several million each week. then transmitted back requested shipment quantities

(a) Legacy process (b) New process

Assortment decisions Assortment decisions

Store inventory Past sales data Requested shipment quantities Past sales data
for each article and size

Store Forecasting
managers model

Inventory Requested shipment quantities Warehouse Inventory Demand Warehouse

in stores, for each article and size inventory in stores forecasts inventory
past sales

Warehouse Optimization
allocation team model

Shipments Shipments

Figure 2: In contrast with the legacy process, the new process relies on formal forecasting and optimization
models to determine weekly shipments to stores (Caro and Gallien 2009).
 Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
74 Interfaces 40(1), pp. 71–84, © 2010 INFORMS

(possibly zero) for each size of each article. A team (1) these forecasts, (2) the inventory of each article
of employees at the warehouse then aggregated and and size remaining both in the warehouse and each
reconciled the requests from all store managers by store, and (3) the assortment decisions. The model
modifying (typically lowering) these requested ship- has shipment quantities as its main decision variables
ment quantities so that the overall quantity shipped and the maximization of global sales as its objective.
for each article and size was feasible in light of the In the sections below, we provide more details on
remaining warehouse inventory. our analytical work, the IT implementation, and the
For many years, the legacy process seems to have implementation and project management that were
effectively supported the relatively small distribution necessary to deploy this new process at Zara.
network for which it had been designed originally.
A key motivation for our project was Zara’s real- Analytical Development
ization in 2005 that the recent growth of its net- As Figure 2(b) illustrates, the analytical development
work to more than 1,000 stores might justify address- of the new process comprised two key steps: the fore-
ing several related improvement opportunities and casting model and the optimization model.
ultimately designing a more scalable process. One The forecasting model generates a prediction of the
issue centered on the incentives of store managers, upcoming weekly demand for each size of each article
whose compensation and career promotion prospects in each store in Zara’s network and essentially relies
are driven significantly by the total sales achieved on the standard methodology of regression analysis.
in their stores. We believe that this caused store We refer the reader to Correa (2007) for further techni-
managers to frequently request quantities exceeding cal details and a complete definition. The most note-
their true needs, particularly when they suspected worthy feature of the implemented forecasting model
that the warehouse might not hold enough inventory may be its high-level structural form as a weighted
of a top-selling article to satisfy all stores. In addi- linear combination of two primary sources of input
tion, store managers are responsible for many tasks data: (1) the objective and centralized time-series data
beyond determining shipment quantities, including of historical sales for the article, and (2) the subjec-
building, sustaining, and managing teams of sev- tive and decentralized shipment request by the store
eral dozen sales associates in environments with manager, converted into a sales prediction by consid-
high employee turnover; thus, they are subject to ering store inventory and existing guidelines about
significant time pressures that compete with their target sales coverage. The weight conferred to the
involvement in determining merchandise replenish- second input source by the least-square fitting pro-
ment requests. Finally, we also believe that the very cedure may thus be interpreted as a credibility mea-
large amount of data for which the warehouse allo- sure (for forecasting purposes) of the store manager’s
cation team was responsible for reviewing (i.e., ship- input. Although the intuitive appeal of this interpreta-
ments of several hundred articles offered in several tion helped Zara to communicate internally about the
sizes to more than 1,000 stores) created significant forecasting model, this ultimate functional form was
time pressures and made it challenging to balance only identified through extensive empirical testing of
inventory allocations manually across stores and arti- many possible predictive equations. In addition, our
cles in a way that would globally maximize sales. team viewed the forecasting model as a modular com-
ponent of the process, in the sense that it can and
should be improved relatively independently from
The Solution the distribution optimization model that we discuss
Figure 2(b) illustrates the structure of the new pro- next. In particular, in addition to continuing to experi-
cess we developed to help Zara compute its weekly ment with other functional forms and additional fore-
store shipments. At a high level, it consists of using casting data sources (e.g., weather), Zara might in
the shipment requests from store managers and past time consider introducing formal incentives for store
historical sales to build demand forecasts. It then uses managers to provide accurate forecasts, adding to its
the following as inputs to an optimization model: more traditional sales-related incentives.
Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
Interfaces 40(1), pp. 71–84, © 2010 INFORMS 75

Expected weekly display policies discussed previously in the Manage-

store sales for
an article (all sizes) rial Challenge section and illustrated in Figure 1, and
Saturation effect
constitutes an important feature of our model: if the
Expected model considered different sizes independently, noth-
demand
(forecast) ing would prevent it from determining that Zara
should ship some units of a given size (e.g., XS) to
a store that would not even display them for sale
during the following week (let alone generate the
sales predicted by the model), because the remain-
Exposure
Stock of article in
ing inventory of a major size (e.g., M) of the article
effect
store at beginning is not sufficient to trigger display at that store. In
of week (all sizes)
other words, the inventory-to-sales function we have
constructed captures the dependencies across sizes
Figure 3: The inventory-to-sales function associated with each store cap-
tures both exposure and saturation effects.
that are introduced by Zara’s policy of only display-
ing for sale the articles with a complete size pro-
file. When considering the aggregate effect across all
The first step of our optimization-model develop- sizes of an article, this gives rise to the exposure effect
ment centered on constructing a predictive model for (Figure 3), whereby the model will (correctly) predict
the expected upcoming weekly sales of all sizes of that no sales will occur in a particular store if that
a given article in a given store, as a function of the store has insufficient available inventory (i.e., does
relevant demand forecasts and the starting level of not include enough different sizes). Incidentally, other
inventory in that store at the beginning of that week. firms that do not employ similar size-based display
We quickly started referring to this predictive model policies might also observe the same effect because of
as the inventory-to-sales function; our motivation for
the self-advertising function of inventory in the store’s
this work was to ultimately support an optimization
sales display area (Smith and Achabal 1998). This fea-
model allowing Zara to predict the impact of ship-
ture is critical from an inventory distribution stand-
ment decisions (affecting the inventory available at
point, because when all sales predictions for different
the beginning of each weekly replenishment period
stores are considered together, the exposure effect will
in each store) on global expected network sales. From
push the optimization model to ship to those stores
a mathematical standpoint, this model resulted from
that can complete a full set of sizes (instead of scat-
the analysis of a stochastic model that considers the
tering some limited inventory over the entire store
sales opportunities of different sizes of an article
as independent Poisson processes. Caro and Gallien network, which could leave many stores below the
(2009) provide additional technical details; however, exposure threshold).
in this paper, we focus on the qualitative and man- The second critical feature of the inventory-to-sales
agerial implications of this function (Figure 3). function is the saturation effect, which reflects the
A first feature of this inventory-to-sales function is decrease of the marginal probability of sale as the
that it considers the entire profile of inventory avail- store receives additional inventory units beyond a cer-
able across all sizes at the beginning of the week. tain point (as highlighted in the upper right area of
In that sense, the representation in Figure 3, which Figure 3). This is a much more classical feature for
suggests a one-dimensional, continuous-input vari- inventory distribution models (Zipkin 2000) that tend
able (the sum of inventory across all sizes), is a to balance the resulting shipments of merchandise
simplification for exposition purposes of the actual across stores, so that each additional unit of available
underlying model (whose input is discrete and multi- inventory is sent on the margin to the store at which
dimensional). This function therefore jointly consid- it has the highest probability of selling. When consid-
ers all sizes simultaneously (as opposed to each size ered jointly, the exposure and saturation effects enable
independently); this is motivated by the inventory the model to select stores that it will bring above the
 Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
76 Interfaces 40(1), pp. 71–84, © 2010 INFORMS

Maximize P ∗ NetworkStoreSales + K ∗ FinalWarehouseStock term that captures expected revenues for the article
Subject to Shipments ≤ InitialWarehouseStock considered over the next week and replenishment
NetworkStoreSales
= Inv-to-Sales (StoreInventory + Shipments)
cycle (denoted as P ∗ NetworkStoreSales in Figure 4),
FinalWarehouseStock = InitialWarehouseStock − Shipments and a second special term equal to the total inputted
value for any inventory remaining in the ware-
Figure 4: Our single-period MIP model formulation involves a trade-off
between sales from the current week and the value of inventory remaining
house after the shipments considered (denoted as K ∗
in the warehouse. FinalWarehouseStock in Figure 4). Although revenues
in the first term are calculated from the known actual
unit selling price P (assumed here to be constant
exposure threshold, and properly balance inventory across stores for expositional simplicity), remaining
between them. warehouse units are valued in the second term with
The last step of our model development work was the aggressiveness factor K, which the user provides
to formulate a mixed-integer program (MIP) embed- and we can thus interpret as the unit value of arti-
ding piecewise-linear approximations of many inde- cles left in the warehouse. A high value of K relative
pendent inventory-to-sales functions, each associated to the store selling price P results in “conservative”
with a store in the network, and allowing the compu- shipments, possibly appropriate shortly after a prod-
tation of shipment quantities that maximize network- uct introduction (when forecast uncertainty is high) or
wide expected sales, subject to inventory availability when the returns and transshipment costs associated
constraints. The piecewise-linear approximations pre- with excessive inventory sent to low-selling stores
serve the essential features of the inventory-to-sales might be particularly high. In contrast, a low relative
function discussed above and make the model solv- value of K results in “aggressive” shipments, perhaps
able with commercial MIP software. This model suitable when forecasts are deemed more reliable and
jointly computes shipment decisions of all sizes of (or) toward the end of the planned shelf life of an
a given article to all stores worldwide in any given article, a time when freeing up some space in the
week; however, for feasibility reasons, it ignores any warehouse for other articles is desirable.
dependencies between shipment decisions of different
articles (i.e., Zara solves an instance of this optimiza- IT Implementation
tion model for each article each week). The IT implementation of the OR models described
Figure 4 shows a high-level representation of the above was relatively challenging, in that it required
single-period MIP formulation, where Inv-to-Sales establishing dynamic access to several large, live
denotes the inventory-to-sales function described databases (store inventory, sales, and warehouse
above; we refer the reader to the appendix and Caro inventory) to compute, under very stringent time con-
and Gallien (2009) for technical details. One of the straints, many decisions that are critical to the com-
most original aspects of this formulation is arguably pany’s operation; in any typical week, Zara now
the use of a control parameter that we called the solves approximately 15,000 instances of our large-
aggressiveness factor (denoted as K in Figure 4). In scale MIP model to distribute six million units of stock
effect, this control is motivated by the fact that the valued at more than E120 million.
model otherwise ignores the important issues of fore- Coauthors Correa and Corredoira developed the
cast uncertainty, the time horizon, and the opportu- core computational application using AMPL and
nity cost of storage space at the warehouse. Because ILOG CPLEX. It involves database queries into Indi-
developing an optimization model that explicitly cap- tex’s legacy IBM AS/400 system, which feeds a
tures these issues presented significant analytical and specific database (SQL Server). From a hardware
data availability challenges, we enabled the user to standpoint, this application is currently executed on
affect the optimization-model outcome according to a two dedicated Dell servers (CPU 3.20 GHz, RAM
procedure that requires some subjective input but is 4 GB, HD 80 GB, OS Windows 2003 Server) located in
specifically designed to consider these issues. Specif- Zara’s IT headquarters in La Coruña, Spain; each han-
ically, our model’s objective includes a standard first dles computations for about half of Zara’s stores. The
Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
Interfaces 40(1), pp. 71–84, © 2010 INFORMS 77

Figure 5: Warehouse allocation employees use a new client application to run the optimization engine, perform
what-if scenario analyses, and visualize and modify the recommended shipments.

implementation also required the in-house develop- virtual private network (VPN); however, a more effi-
ment of a client application, which we distributed on cient node-based infrastructure is being deployed.
the PCs of the warehouse allocation team’s approx-
Implementation and Project Management
imately 60 employees. This second application was
In 2005, when the first project-related discussions
developed using Visual FoxPro and provides an inter- started, implementing OR models on a large scale at
face that allows these employees to request additional Zara to support core business decisions, which would
runs of the core computational application to perform affect the entire company’s success, seemed like a
what-if scenario analyses, visualize, and manually daunting task. Zara is a fashion company with a cul-
modify any output of the optimization model, and ture that strongly favors human intuition, vision, and
communicate their chosen solutions to the existing judgment (as opposed to analytical methods) for deci-
warehouse control systems; thus, they are effectively sion making. It has a history of success in showing
that these subjective or nonquantitative approaches
implementing the physical picking, sorting, packing,
can pay off when applied to many of the key decisions
and truck-loading operations corresponding to the
it faces (e.g., design of clothes). OR awareness within
shipments determined by the warehouse allocation
Zara essentially did not exist when the authors started
team using the decision support system. Figure 5 collaborating on this project. Bringing Zara to its cur-
shows a screen snapshot of this client application. rent state, in which many employees and some key
Finally, the data-communications infrastructure that executives trust OR methods to determine key opera-
supports these applications relies, to date, on Inditex’s tional decisions, required a substantial investment in
 Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
78 Interfaces 40(1), pp. 71–84, © 2010 INFORMS

communication and education by all the authors, and description of how the forecasting model uses this
careful thinking about the design and management of information). Although directly requesting demand
this implementation project. forecasts from store managers might improve overall
In particular, we believe that some design choices forecast accuracy in the long run, the practical difficul-
were critical to our success in this environment. One ties and resource requirements associated with man-
was the role of the warehouse allocation team in the aging such a significant operational change, which
new process. We clearly established and communi- would affect a global network of more than 1,500
cated very early on that the forecasting and optimiza- stores, prompted the team not to undertake this effort
tion models were not meant to replace that team, but within the targeted project timeline. We note, how-
rather to assist its members in performing their tasks ever, that Zara might still leverage this opportunity to
more effectively. This orientation induced several improve the new process in the future and that more
technical choices, such as developing the distributed general forecasting-procedure improvements can be
client application to allow the warehouse employ- performed in a modular way without affecting the
ees to perform what-if analysis scenarios defined, for optimization model’s structure.
example, by the aggressiveness factor K discussed We believe that several aspects of managing this
in the Analytical Development section, and to manu- project were material success factors. The first was
ally and freely modify all the individual shipments the project team’s organizational structure. Specifi-
computed by the optimization model before their cally, our work began with collaboration between aca-
physical execution. The academic project team mem- demic researchers (Caro and Gallien) and industry
bers were initially uncomfortable with this last fea- practitioners (Ramos, García, and Montes). Although
ture, because of methodological concerns linked to such collaborations generally have the potential to be
measuring impact for the optimization model output. powerful and mutually enriching, they also present
However, we realized that this choice was essential intrinsic challenges because of possible differences
to secure the warehouse allocation team’s involve- in objectives, incentives, and time horizons. In our
ment and to allow us to benefit from the team’s con- case, we observed that the jointly supervised six-
siderable knowledge of the managerial challenge at month internship at Zara of (then) graduate student
stake; this knowledge proved invaluable to us as we Correa, who was enrolled in MIT’s LFM program,
formulated the optimization model. In addition, we provided an effective coordination mechanism and
recognize now that these concerns were unfounded; liaison between the two parts of the team. It also
after a live trial and debugging period of a few served as a catalyst for the model development work
weeks, the warehouse allocation team members rarely performed in preparation for that internship, which
(if ever) modified the optimization model’s output was to focus on implementing these models.
directly (as opposed to changing the model’s input Another important catalyst was the live pilot exper-
and control parameters). In other words, allowing iment of the new process, which the entire team
them to have total flexibility and control over the out- planned as a critical project requirement and was
put was essential to develop their internal confidence designed with three goals: (1) provide a convinc-
in the model, although they ultimately rarely used ing proof of feasibility for the new OR-based pro-
this feature. cess, (2) help identify and act upon improvement
From an implementation standpoint, a second opportunities for the new process before its full-scale
important process feature was our choice to leave deployment, and (3) support a quantitative impact
the information interface with stores as it was in the assessment. Partly because the live pilot would affect
legacy process. That is, the store manager input that the shipments of real merchandise to several hundred
the forecasting model required in the new process stores and would also involve a friendly competition
was the requested shipment quantities that the man- with the legacy process (the Impact section provides
agers were already providing as part of the legacy details), it helped greatly to define priorities and focus
process (the Legacy Process and Analytical Develop- the energies of all team members. It also helped to
ment sections provide background information and a overcome the cultural barriers mentioned above and
Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
Interfaces 40(1), pp. 71–84, © 2010 INFORMS 79

Zara assigns Deployment

full-time IT completed
Zara approves engineer in Zaragoza
LFM internship
J. Correa’s LFM
internship Deployment
completed
First in Arteixo
meeting at Model Software
Zara HQ development Forecast model prototype Process
and test development development Live pilot deployment

Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Jun

2005 2006 2007

Figure 6: The project involved five main phases spanning almost two years.

generate buy-in from important stakeholders within excluding the labor costs of all contributing employ-
the company when the team subsequently planned ees not specifically hired for this project.
and executed the full-scale deployment of the new
process. Impact
The complete development and implementation
cycle for this work spanned just under two years Measurement Methodology
(Figure 6), and its corresponding total cost to Zara The methodology to estimate the implementation im-
is estimated to range between $150,000 and $250,000, pact involved a live pilot implementation experiment

Arteixo stores
Test article
Relative Control article
sales
comparison
Zaragoza stores

Test article
Optimization model and Shipments of control
new process used to article determined
determine shipments with legacy process
for test article
Relative
Shipments of both sales
Arteixo test and control comparison
warehouse articles are determined
with legacy process

Zaragoza
warehouse
Control article

Figure 7: The experimental design of the live pilot involves two dimensions of control for each test article; the
illustration shows 1 of the 10 test articles and its paired control article.
 Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
80 Interfaces 40(1), pp. 71–84, © 2010 INFORMS

using two dimensions of control (Figure 7). Using the work could only be attributed to estimation errors, as
optimization model, we distributed a test group of opposed to the use of our new OR-based process.
10 articles to all stores that exclusively receive ship-
ments from Zara’s first primary warehouse in Arteixo, Financial Impact
Spain (this represents approximately half of all Zara Figure 8 summarizes the main results of the con-
stores worldwide). Simultaneously, we distributed a trolled live pilot experiment described above. These
subset of 10 “twin” control articles, which we deter- results are conclusive: The control-adjusted relative
mined through a careful pairwise matching with the sales impact in Arteixo is positive for every article,
test group, to the same stores using the legacy manual with a mean across articles of 4.1 percent (median
process. This matching procedure enabled us to do 4.2 percent); the corresponding estimation error cal-
a relative comparison between the test and the con- culated using data from Zaragoza is centered around
trol group and thus supported estimating the specific
zero (mean and median across articles are 0.7 and
impact of the new model and eliminated the impact
−0!6 percent, respectively). We also note here that
of any other factors external to the model (which
the variability around the means that we see across
would affect both the test and control groups); this
different articles in Figure 8 is easily explained by
constitutes the first dimension of control. The second
the measurement noise introduced for each individ-
dimension of control exploits the fact that the other
ual article by the pairwise matching procedure and
Zara stores worldwide receive only shipments from
its second main warehouse in Zaragoza, Spain. For by possible forecasting errors. Therefore, focusing on
this second and relatively independent network of the averages of these measurements across articles,
stores, we determined shipments of both test and con- which are indicative of the model’s overall impact on
trol group articles using the legacy manual process. the entire range of Zara’s product offerings, is appro-
We selected this experimental design to enable us to priate. When (conservatively) subtracting the esti-
estimate the error associated with the control-based mated average experimental error obtained from the
impact-estimation methodology described above. For Zaragoza data, these results show that the new OR-
Zaragoza stores, we distributed both test and control based process increases sales during the selling season
group articles using the same procedure; any mea- by 3–4 percent, as seen in Figure 8. This impact on
sured impact for this second half of the store net- revenue is easily explained by the model’s ability,

Arteixo
Zaragoza

Each point on this graph

represents the difference in
relative change of sales during the
live pilot between a test article
Difference in relative
and its paired control article
change of sales
between test and
–10.0 –7.5 – 5.0 – 2.5 0.0 2.5 5.0 7.5 10.0 control articles (%)

Figure 8: The results of the live pilot experiment suggest that the new process increases sales by approximately
3–4 percent (adapted from Caro and Gallien 2009).
Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
Interfaces 40(1), pp. 71–84, © 2010 INFORMS 81

relative to the legacy process, to determine that Zara vision, and judgment (as opposed to analytical meth-
should move excessive inventory away from low- ods) when making decisions. This is partly because the
selling stores where it is not needed, send it to high- typical background of most fashion-industry employ-
performing stores in which it reduces the number of ees is not quantitative and perhaps also because Zara
sales missed because of stock-outs, and ship all sizes owes much of its success to the unique intuition of its
of an article to each store in a concerted manner, like- founder (Ghemawat and Nueno 2003). We doubt that
wise sending inventory of specific sizes only to stores Zara will ever use OR models to help with several of
in which it is likely to sell. its key challenges, including anticipating volatile mar-
From a financial standpoint, if Zara had used the ket trends, recruiting top designers, and creating fash-
optimization model for all of 2007 and 2008, this rela- ionable clothes. Moreover, it is not clear to us that the
tive sales increase would have implied approximately company should. However, the work presented here
$310 million (2007) and $353 million (2008) in addi- did prompt a realization by many of Zara’s key execu-
tional revenue or $37.2 million (2007) and $42.4 mil- tives and employees that for processes involving large
lion (2008) in additional net income, with both mea- amounts of quantitative data, well-designed OR mod-
sures of impact predicted to continue growing at a els will lead to better performance and more scalable
rate of 10 percent per annum in subsequent years. operations. Because of this work, Zara has initiated
However, the model’s full-scale deployment started two additional major OR implementation projects on
in late 2006 and was completed in June of 2007; purchasing and pricing. It is now also actively seek-
therefore, the estimated, actual realized impact of the ing to recruit graduates with strong OR backgrounds
model in 2007 is instead approximately $233 million and to become a corporate partner with MIT’s LFM
in additional revenue or $28 million in additional net Program. In addition, the Inditex Group is planning
income (that is, 75 percent of the previous figures for to deploy the OR-based inventory distribution pro-
2007). The companion paper, Caro and Gallien (2009), cess described in this paper in some of its other retail
presents some quantitative evidence that the model chains, such as Massimo Dutti.
also reduced the transshipments between stores and
increased the time spent by articles on display within Summary of Contributions to
their life cycle; however, the financial implications of
OR Practice
these observations are harder to estimate.
To the best of our knowledge, this work constitutes
Organizational Impact the first reported application of OR to the fast-fashion
Zara was able to maintain its warehouse inventory retail-business strategy, as adopted by companies that
allocation team at its early 2007 staffing level of include Zara, H&M, and Mango. Under this strategy,
approximately 60 individuals worldwide, although the life cycle of merchandise in these stores spans
it was initially planning on expanding that team only a small fraction of a selling season (e.g., five
because of its projected 10–12 percent sales growth to six weeks for Zara). Therefore, their store assort-
per annum. More importantly, the optimization model ments turn over much more frequently than those of
has had a significant impact on the daily lives of these traditional retailers, customers find frequent store vis-
employees by enriching their professional responsi- its to be more appealing, and the warehouse inven-
bilities: all team members have become enthusiastic tory available for distribution to these stores tends to
users of the new tool, gratefully seeing their respon- be scarcer than for traditional retailers, making the
sibilities shift from repetitive manual data entry to inventory allocation problem more difficult.
exception handling, scenario analysis, and process This work also appears to constitute the first
improvement. described implementation of an inventory distribu-
tion model for an apparel retailer (fast-fashion or
Cultural Impact other), which specifically captures the dependencies
This project has also had some cultural impact at across sizes introduced by store inventory display
Zara, a company where many favor human intuition, policies. Specifically, many retailers, including Zara,
 Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
82 Interfaces 40(1), pp. 71–84, © 2010 INFORMS

remove some articles from display and put them research universities during this project), and it
into the store’s backroom whenever the combination has significantly affected their research and teaching
of their sizes still available is not complete enough. activities. In particular, this OR implementation has
Although many retailers replenish their stores for already been the object of several masters, executive
each size independently (sometimes leading to use- education, and MBA course sessions taught at UCLA,
less shipments from a sales standpoint if major sizes MIT, and Columbia University. All team members
are missing) or ship predetermined size bundles also learned many lessons, several of which were ini-
(making it difficult to balance the inventory avail- tially counterintuitive, about implementing OR. For
able across sizes in stores), our model considers the example, from an academic standpoint, designing and
entire inventory profile of all sizes offered and com- studying OR models that capture all the relevant key
putes coordinated shipment quantities for all these input data (including forecast uncertainty and stor-
sizes simultaneously; the Analytical Development sec- age opportunity cost in our case) might seem impor-
tion provides more detail. tant. In practice, however, our experience suggests
Finally, we emphasize that although our impact- that implementing such “comprehensive” models to
measurement methodology based on a controlled control large-scale core processes in time-sensitive
experiment is fairly common in other disciplines (e.g., environments could present insurmountable IT and
medicine and social sciences), its application to OR data-availability challenges; thus, a hybrid approach
practice is noteworthy, because the impact of pub- combining a simpler but robust optimization model
licly described OR practice work is predominantly with carefully selected user inputs might be more
estimated through more questionable “before versus effective.
after” comparisons that completely ignore the many In closing, we note that this paper presents a suc-
other factors that are not related to OR, but could also cessful and high-impact application of OR techniques
be affecting differences in performance observed in in a fashion-retailing environment, which is both
the “after” period. highly visible to the general public and not currently
perceived as a traditional application area for this dis-
cipline. We are hopeful that our work will contribute
Concluding Remarks to increasing OR awareness and help improve percep-
Because of Zara’s openness to the academic pub- tions of its applicability.
lication of this work, the OR models presented
might also impact other companies. Although the Appendix
inventory distribution model discussed above would
Single-Store Inventory-to-Sales Model
clearly require some IT implementation and adap- Consider an article offered in a set of sizes S = S + ∪ S − ,
tation work, it seems particularly applicable to the where S + denotes the major sizes (e.g., {S, M, L}) and S −
many apparel retailers facing the challenge of coor- the minor sizes (e.g., {XS, XL}). Sale opportunities for each
dinating shipment decisions across different sizes size s ∈ S are assumed to be independent across sizes and
follow a Poisson process with rate "s and cumulative count-
because of store display policies. A simpler version of
ing measure {Ns #t$% t ≥ 0}, where t denotes the time elapsed
this model, which would not capture sizes but would since the last replenishment (i.e., Ns #t$ is the random num-
capture the notion of required store-exposure inven- ber of sale opportunities for size s that occurred between
tory, also seems applicable to most of the retail indus- 0 and t). Let qs represent the inventory level of size s imme-
try beyond the apparel segment and would likely diately after replenishment at time 0; the virtual stockout
time &s #qs $ can be defined for every size s ∈ S as
constitute a significant improvement relative to the
'
pervasive non-OR-based methods and simple heuris-
! "
&s #qs $ = inf t ≥ 0( Ns #t$ = qs !
tics that many firms use when distributing scarce
Likewise, the earliest time at which one of the major sizes
inventory to a network of stores (e.g., proportional
runs out, from an initial profile q of inventory across sizes
rationing). and assuming no replenishment occurs, can be expressed as
Our work has also been very formative for its
'
academic authors (both were junior faculty at major &S + #q$ = min &s #qs $!
s∈S +
Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
Interfaces 40(1), pp. 71–84, © 2010 INFORMS 83

As we described above, all inventory is removed from cus- Decision Variables

tomer view as soon as one of the major sizes runs out at any xsj ∈ N : Shipment quantity of each size s ∈ S to each store
point between successive replenishments. Under that pol- j ∈ J for the current replenishment period;
icy, the (random) total number of sales in a replenishment zj : The approximate expected sales across all sizes in each
period can be expressed as store j for the current period under consideration;
' #
G#q$ = Ns #&S + ∧ T $ +
#
Ns #&S + ∪)s* ∧ T $% yj : Secondary variables representing the term
s∈S + s∈S −
! "
where T > 0 denotes the time between consecutive replen- min ai #"s′ j $#Is′ j + xs′ j − i$ + bi #"s′ j $ / and
' s∈S + % i∈N #"sj$
ishments (one week for Zara) and a ∧ b = min(a% b). Apply-
ing Doob’s optional sampling theorem, the expectation
' vsj : Secondary variables representing the term
g#q$ = E+G#q$, can be expressed as
# ! "
g#q$ = "S + E+&S + ∧T ,+ "s E+&S + ∪)s* ∧T ,% min ai #"s′ j $#Is′ j + xs′ j − i$ + bi #"s′ j $ !
s∈S − s ′ ∈S + ∪)s*% i∈N #"s ′ j$

' #
where "S + = "s !
s∈S +

For any subset of sizes D ⊂ S , we apply Jensen’s approxi-

mation, E+&S + ∧ T , ≈ mins∈D E+&s ∧ T ,, and observe that Objective
$ $ $
# Max j∈J Pj zj + K# s∈S #Ws − j∈J xsj $$.
E+&s ∧ T , = -#k% "s T $/"s . #k$
k=1!!qs

Constraints
! "
= min ai #"s $#qs − i$ + bi #"s $ %
i∈N $
j∈J xsj ≤ Ws for all s ∈ S (warehouse inventory availabil-
' ' '
where ak #"s $ = -#k% "s T $/"s . #k$ and a+ #"s $ = 0% bi #"s $ = ity constraint);
' '
$ $
$ zj ≤ # s∈S + "sj $yj + s∈S − "sj vsj for all j ∈ J (primary
k=1!!i−1 ak #"s $ for i ≥ 1% b0 #"s $ = 0% and b+ #"s $ = T , and
. and - are the Gamma function and the lower incomplete inventory-to-sales function implementation constraint);
Gamma function, respectively. Our next approximation con- yj ≤ ai #"sj $#Isj + xsj − i$ + bi #"sj $ for all j ∈ J , s ∈ S + , and
sists of computing only the minimum on the right side of i ∈ N #"sj $ (secondary inventory-to-sales function implemen-
the previous equality over the small finite subsets N #"s $ tation constraint);
defined as vsj ≤ )ai #"sj $#Isj + xsj − i$ + bi #"sj $* for all j ∈ J , s ∈ S − , i ∈
' ! "
N #"s $ = i ∈ N ∪ )+*( bi #"s $ ≈ 0% 0!3T % 0!6T % 0!8T % 0!9T % T % N #"sj $ (secondary inventory-to-sales function implementa-
tion constraint);
which are straightforward to compute numerically. Appro- vsj ≤ yj for all j ∈ J , s ∈ S − (secondary inventory-to-sales
priate substitutions yield the final expression: function implementation constraint); and
! "
g#q$ ≈ "S + min ai #"s $#qs − i$ + bi #"s $ xsj ∈ N ; zj % yj ≥ 0; vsj ≥ 0 (nonnegativity and integer con-
s∈S + % i∈N #"s$
# ! " straints).
+ "s min ai #"s′ $#qs′ − i$ + bi #"s′ $ !
s ′ ∈S + ∪)s*% i∈N #"s ′ $
s∈S −

Caro and Gallien (2009) provide a complete discussion.

Acknowledgments
We thank the Zara Distribution Team—in particular, Juan
Network Sales-Optimization Model Quintela, Jesús González, and Iván Escudero; the Inditex
Communications Department—in particular, Jesús Echevar-
Input Data ria and Amaya Guillermo; Anthony Placet from the MIT
Set of sizes: S = S + ∪ S − partitioned into major sizes S + Sloan Office of Communications; Donald Rosenfield and
and regular sizes S − (index s); Jonathan Griffith from the MIT Leaders For Manufacturing
Set of stores: J (index j); Program; all the volunteers who made the 2009 Franz Edel-
Ws : Inventory of size s available in the warehouse;
man Competition possible, and in particular Coaches Layek
Isj : Inventory of size s available in store j;
Abdel-Malek, Ananth Iyer, Yoshi Ikura, and Chair Srinivas
Pj : Selling price in store j;
K: Aggressiveness factor (value of inventory remaining Bollapragada; the 2009 Franz Edelman Competition spon-
in the warehouse after the current shipments); sors, Georgia Institute of Technology, Procter & Gamble,
"sj : Demand rate for size s in store j; and and SAS. F. Caro and J. Gallien would also like to thank
N #"sj $: Approximation set for size s in the inventory-to- faculty colleagues at the UCLA Anderson School of Man-
sales function approximation for store j. agement and MIT Sloan School of Management.
 Caro et al.: Zara Uses Operations Research to Reengineer Its Global Distribution Process
84 Interfaces 40(1), pp. 71–84, © 2010 INFORMS

References Ghemawat, P., J. L. Nueno. 2003. ZARA: Fast fashion. Harvard

Business School Multimedia Case 9-703-416, Harvard Univer-
Caro, F., J. Gallien. 2009. Inventory management of a fast-fashion sity, Boston.
retail network. Oper. Res., ePub ahead of print August 12, http://
or.journal.informs.org/cgi/content/abstract/opre.1090.0698v1. Helm, B. 2008. Best global brands. BusinessWeek (September 18) 52.
Correa, J. 2007. Optimization of a fast-response distribution net- Smith, S. A., D. D. Achabal. 1998. Clearance pricing and in-
work. Unpublished master’s thesis, Massachusetts Institute of ventory policies for retail chains. Management Sci. 44(3)
Technology, Cambridge. 285–300.
Fraiman, N., M. Singh, L. Arrington, C. Paris. 2002. Zara. Columbia Zipkin, P. H. 2000. Foundations of Inventory Management.
Business School case, Columbia University, New York. McGraw-Hill, New York.