Zara Uses Operations Research to Reengineer Its Global Distribution Process

Miguel Díaz, CFO

Javier García

José Manuel Corredoira

José Antonio Ramos

Juan Correaformerly with ZaraJosé Manuel Corredoira

Marcos MontesZara, Inditex Group S.A.

formerly with Zara

Felipe CaroUCLA Anderson School of Management

Jérémie GallienMIT Sloan School of Management

Customer feedbackand fashion trendsDistributionDistributionDistributionDistribution

Stores

Global Warehouses Design Team

Zara’s Continuous Business Cycle

Production planningand specifications

Productionand delivery Suppliers

Global Warehouses Design Team

Legacy Distribution Process (2006 and Before)

past sales data

storemanagers

store inventory

assortment decisions

warehousedistribution team

shipments

inventory

in stores,

past sales

warehouse

inventory

requested shipment

quantities for each

article and size

Problems with Legacy Process

• Store incentives andlocal optimization

• Warehouse resourcesand guidelines

lastweek

store

weeksales

stocklevel

order

New OR-Based Distribution Process

past sales data

storemanagers

requested shipment

quantities for each

inventory

in stores,

warehouse

inventory

store inventory

requested shipment

quantities for each

inventory

in stores

warehouse

inventory

assortment decisions

demand

forecasts

past sales data

forecastingmodel

assortment decisions

quantities for each

article and size

warehousedistribution team

shipments

in stores,

past sales

inventoryquantities for each

article and size

optimizationmodel

shipments

in stores inventory

Legacy Process New Process

forecasts

The Optimization Model

For every article:

Max Σstore j Price(j) x Sales(j)

Subject to:

Expected revenue for upcoming week Value of leftover inventory

+ Input_Value x Σsize s Leftover_WH_Inventory(s)

Subject to:

Shipment(j,s) is integer

Σstore j Shipment(j,s) ≤ Existing_WH_Inventory(s)

Sales(j) = Function[ Store_Inventory(j,s) + Shipment(j,s) , Forecast(j,s) ]

Leftover_WH_Inventory(s) = Existing_WH_Inventory(s) - Σstore j Shipment(j,s)Inventory-to-sales function

Inventory availability constraint

Leftover inventory calculation

Inventory-To-Sales Function

Expected store salesfor article that week

Expected demand(Forecast)

Saturation Effect

Stock of articlein store atbeginning of week

Exposure Effect

Exposure Effect and Size Display Policy

remainingsizes

action

Article with 4 Sizes(Small, Medium, Large, Extra Large)

S M L XL Keep on display

M L XL Keep on display

S M L Keep on display

M L Keep on display

Entire article is removed to backroom when major size Medium or Largestocks out

M L Keep on display

S M XL Move to backroomS M XL Move to backroom

S L XL Move to backroom

S M Move to backroom

L XL Move to backroom

S L Move to backroom

M XL Move to backroom

S XL Move to backroom

S Move to backroom

M Move to backroom

L Move to backroom

XL Move to backroom

Stochastic Store Display Model

Sizes

Poisson processarrival rates

obtained fromdemand forecasts

Starting inventory xs of each size s:

x = Store_Inventory(s) + Shipment(s)

Major sizes Regular sizes

InventoryCustomer Arrival Process

SMλSM

MλM

XLλXL

LλL (set A)

•Stochastic stopping times:

–The time when a size stocks out:

–The time when the first major size stocks out:

•Linear approximation of expected sales function:

xs = Store_Inventory(s) + Shipment(s)

= Function[ Store_Inventory(s) + Shipment(s) , Forecast(s) ]

Mathematical Model FormulationExpected revenue for

upcoming weekValue of leftover

inventory

Max

Subject to:

Approximateinventory-to-sales

function

Inventory availabilityconstraint

Pilot Test: Objectives

• When: 2006 Fall/Winter season

• Objectives:

1. Prove concept feasibility through actual implementationimplementation

2. Refine software interface and model features based on feedback from field users

3. Estimate the model’s specific impact on inventory distribution

Pilot Test: Matching Procedure

control article

• 10 test articles that represent Zara’s assortment

• Each test article was matched with a control article

test article

• Matching criteria:

− same type of article

− same life-cycle

− similar prior performance

Pilot Test: Experiment Design

Zaragoza stores

Arteixo stores

Optimization modelcomputes shipments

test article control article

Shipments determinedwith legacy process

Performancecomparison

test article

Arteixowarehouse

Zaragozawarehouse

computes shipments

control article

Shipments determinedwith legacy process

Performancecomparison

ControlTest MM ′−′

Pilot Test: Arteixo Warehouse

test article

legacy process for test and control articles

PILOT: new process for test articlelegacy process for control article

Arteixo

ControlTest MM −

Life-cycle End

Life-cycle Start

article

control article

Measuring Point

Pilot Starts

Pilot Test: Zaragoza Warehouse

test article

legacy process for test and control articles

legacy process for test and control articles

Zaragoza

ControlTest MM ′−′

ControlTest MM −

article

control article

Life-cycle End

Life-cycle Start

Measuring Point

TestTest MM −′

ControlControl MM −′

1. Remove anything that happened prior to the pilot

2. Eliminate factors that are common to test and control

for test articles

for control articles

Pilot Test: Impact Estimation Methodology

2. Eliminate factors that are common to test and control articles (1st dimension of control)

• Arteixo: impact of the new process

• Zaragoza: measurement of impact estimation error (2nd dimension of control)

TestTest MM −′

ControlControl MM −′

1. Remove anything that happened prior to the pilot

2. Eliminate factors that are common to test and control

for test articles

for control articles

Pilot Test: Impact Estimation Methodology

)()( ControlControlTestTest MMMM −′−−′

2. Eliminate factors that are common to test and control articles (1st dimension of control)

• Arteixo: impact of the new process

• Zaragoza: measurement of impact estimation error

(2nd dimension of control)

Pilot Test: Increase in Sales

Sales

Arteixo

Zaragoza

-10,0% -7,5% -5,0% -2,5% 0,0% 2,5% 5,0% 7,5% 10,0%

SalesDifference

Pilot Test: Increase in Sales

Sales

Arteixo

Zaragoza

-10,0% -7,5% -5,0% -2,5% 0,0% 2,5% 5,0% 7,5% 10,0%

SalesDifference

Pilot Test: Increase in Sales

Sales

Arteixo

Zaragoza

0.7 % Average Error

-10,0% -7,5% -5,0% -2,5% 0,0% 2,5% 5,0% 7,5% 10,0%

SalesDifference

0.7 % Average Error

Pilot Test: Increase in Sales

Sales

Arteixo

Zaragoza

0.7 % Average Error

4.1 % Average Increase

-10,0% -7,5% -5,0% -2,5% 0,0% 2,5% 5,0% 7,5% 10,0%

SalesDifference

0.7 % Average Error

Pilot Test: Increase in Sales

Arteixo

Zaragoza4.1 % Average Increase

= 3 to 4% Increase in Sales–

-10,0% -7,5% -5,0% -2,5% 0,0% 2,5% 5,0% 7,5% 10,0%

SalesDifference

0.7 % Average Error

Realized Financial Impact

3 to 4% Increase in Sales

Realized Financial Impact

3 to 4% Increase in Sales

2007

Additional

Net Income

Additional

Revenue$233 M

$28 M

Realized Financial Impact

3 to 4% Increase in Sales

20082007

$353 M

$42 MAdditional

Net Income

Additional

Revenue$233 M

$28 M

storesdays

storeperdisplayondays

×

Operational Impact

• Display Cover Ratio

+3.5%

shippedunits

returnedoredtransshippunits

• Transshipments and Returns Ratio

-19%

Conclusion

• Technical OR achievements:– First application of OR to fast fashion

– Tight MIP formulation of complex stochastic problem

– Distributed IT implementation with many real-time users

– Large controlled field experiment to estimate impact

• Business impact:• Business impact:– Substantial measurable increase in sales by 3-4% � $350M

– Shelf exposure time increased by 3.5%

– Transshipments and returns reduced by 19%

– Scalable process without major capital investment

• Human impact: Positive transformation of sixty employees’ daily lives