You are on page 1of 22

This article was downloaded by: [115.113.11.

143] On: 29 January 2017, At: 09:22


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

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

Using Analytics to Enhance a Food Retailers Shelf-Space


Management
Teresa Bianchi-Aguiar, Elsa Silva, Luis Guimares, Maria Antnia Carravilla, Jos F. Oliveira,
Joo Gnther Amaral, Jorge Liz, Srgio Lapela

To cite this article:


Teresa Bianchi-Aguiar, Elsa Silva, Luis Guimares, Maria Antnia Carravilla, Jos F. Oliveira, Joo Gnther Amaral, Jorge Liz,
Srgio Lapela (2016) Using Analytics to Enhance a Food Retailers Shelf-Space Management. Interfaces 46(5):424-444. http://
dx.doi.org/10.1287/inte.2016.0859

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

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

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

Copyright 2016, INFORMS

Please scroll down for articleit is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
Vol. 46, No. 5, SeptemberOctober 2016, pp. 424444
ISSN 0092-2102 (print) ISSN 1526-551X (online) http://dx.doi.org/10.1287/inte.2016.0859
2016 INFORMS

Using Analytics to Enhance a Food Retailers


Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

Shelf-Space Management

Teresa Bianchi-Aguiar, Elsa Silva, Luis Guimares, Maria Antnia Carravilla, Jos F. Oliveira
INESC TEC and Faculty of Engineering, University of Porto, 4200-465 Porto, Portugal
{mtbaguiar@fe.up.pt, elsamsilva@fe.up.pt, guimaraes.luis@fe.up.pt, mac@fe.up.pt, jfo@fe.up.pt}

Joo Gnther Amaral, Jorge Liz, Srgio Lapela


Sonae MC, 4471-909 Maia, Portugal
{jnamaral@sonaemc.com, jorgeliz@sonaemc.com, smlapela@sonaemc.com}

This paper describes the results of our collaboration with the leading Portuguese food retailer to address the
shelf-space planning problem of allocating products to shop-floor shelves. Our challenge was to introduce
analytical methods into the shelf-space planning process to improve the return on space and automate a process
heavily dependent on the experience of the retailers space managers. This led to the creation of GAP, a decision
support system that the companys space-management team uses daily. We developed a modular operations
research approach that systematically applies mathematical programming models and heuristics to determine
the best layout of products on the shelves. GAP combines its analytical strength with an ability to incorporate
different types of merchandising rules to balance the tradeoff between optimization and customization.
Keywords: retail operations; shelf-space allocation; mathematical programming; MIP-based heuristics.

average sales areas of 9,000 m2 (96,875 square feet),


S onae MC, which was ranked as the fourth largest
company in Portugal in 2014 and had annual
sales of 3.33 billion euros, operates a food retail busi-
offering an extensive and varied range of products and
services, and in prime locations. Sonae MC has a total
ness within Portugal. It is one of the core businesses sales area of 595,000 m2 (6,404,527 square feet) and its
of the Sonae Group, which also operates in other strategy is to grow its convenience channel and to look
areas such as specialized retailing (e.g., sports goods, for international growth opportunities.
fashion, and electronics), shopping centers, and Sonae MC is aware of the impact in-store planning
telecommunications. has on customer satisfaction, sales effectiveness, and
Its Continente brand is the countrys leading food re- operations efficiency. In particular, it believes that a
tailer, and has been considered one of the most trusted well-planned and innovative product organization on
brands in Portugal for the past 13 years. Sonae MC its shelves can lead to higher visibility, increased con-
launched Portugals first hypermarket in 1985 and is sumer awareness, and increased demand for its prod-
a benchmark in the Portuguese market. Today it has ucts, in addition to reduced inventory holding and
a network of 478 stores (with an additional 162 stores handling costs. However, the short product life cycles,
under franchising) covering the entire country. Stores increasing number of available products, and continu-
adopt one of the following three formats: Continente ally growing number of stores have led to an ongoing
Bom Diaconvenience stores with an average sales need to review shelf-space planning, an increasingly
area of 800 square meters (m2 ) (8,611 square feet); Con- challenging process for the company.
tinente Modelosupermarkets averaging 2,000 m2 This paper describes the development, implementa-
(21,528 square feet) and located in medium-sized pop- tion, and impact of an operations research (OR)-based
ulation centers; and Continentehypermarkets with approach to planning the allocation of products on
424
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
Interfaces 46(5), pp. 424444, 2016 INFORMS 425

Sonae MCs shelves. This approach is the result Shelf-Space Management at Sonae MC
of a collaboration between the companys informa- The primary objective of retailing is to bridge the
tion systems and innovation department (ISI) and gap between the point of production and the point
its space-planning department, and a group of re-
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

of sales, a goal that stresses the role of logistics and


searchers from the industrial engineering and man-
operations in this industry. Sonae MC has a central-
agement department of the faculty of engineering of
ized operations management department, which is
the University of Porto (FEUP).
responsible for planning all operations for the com-
Prior to this work, Sonae MCs shelf-space planning
panys stores nationwide. The space-planning depart-
process was time consuming, relied on an empirical
ment, as its name implies, is responsible for managing
use of space elasticities, lacked formal performance
the space available at the stores, an activity that has
evaluation criteria, and depended heavily on the
two main levels: (1) a macro-space planning level that
experience of its space managers. Our challenge con-
defines, on a long-term basis, the layout of the stores,
sisted of incorporating analytical methods into the
which are divided by categories (i.e., classes of prod-
practice to automate the process, improve the return
ucts, such as yogurt, milk, and beverages, that have
on space, and reduce stockouts and inventory costs,
shared characteristics), and (2) a shelf-space (or micro-
while questioning (but not disrupting) the companys
space) planning level that defines, for each category,
policies. Guided by these objectives, we defined three
the products placement on the shelves. Shelf-space
axes for the project: process automation, space opti-
planning is a midterm activity that updates shelf-
mization, and display standardization (i.e., organizing
space plans approximately two to three times a year
all stores in a similar manner).
We labeled our resulting decision support system for more than 300 categories. This activity requires 23
(DSS) as automatic generation of planograms and full-time space managers.
use the acronym GAP. We developed GAP on top The traditional shelf-space planning tool is a plano-
of a modular architecture with innovative, tailor- gram, which is a virtual representation of the shelves,
made mathematical programming models embedded showing exactly where each product should be phys-
in matheuristics to ensure its efficiency. In addition to ically displayed and the inventory that each shelf
its analytical advances, the DSS permits the incorpo- should hold. A planogram includes a variety of infor-
ration of various types of allocation rules (i.e., mer- mation that must be planned carefully: examples
chandising rules), allowing space managers to control include the location of the products, the number
the entire process, with a tradeoff between customiza- of facings (i.e., visible items), the number of items
tion with optimization. Sonae MCs space managers stacked behind and above each facing, the packaging
now use GAP daily to automatically generate shelf- style, and the orientations (front, side, back, top). In
space plans. addition to the most commonly used shelves, stores
We structured the remainder of this paper as fol- also have other fixture types, such as chests, pallets,
lows. We start by describing shelf-space management and pegboards (i.e., bars with steel rods that pro-
at Sonae MC from a generic perspective, and from trude to hold peggable products, such as pens and
both practical and theoretical points of view. We then pencils). Moreover, planograms are physically con-
present GAP, including a general discussion of its structed of segments that are stacked together to form
analytical approach and a description of the DSS. an aisle. Each segment has its own shelves: these can
We discuss some details of the projects develop- be placed, aligned with the shelves of the other seg-
ment, which were critical for its success, and ana- ments, or placed with a different orientation, forming
lyze the projects impact. We end with brief conclud- irregular planograms. Figure 1 shows some examples.
ing remarks, emphasizing that we are presenting a At Sonae MC, planograms follow a complex struc-
generic approach that is suitable for use by other ture of merchandising rules that try to reflect con-
retail companies. This is a practice-oriented paper in sumer buying behavior and the strategy of the
which we omitted many details for simplicity and company (and of its suppliers) for the various cate-
brevity. Therefore, throughout the text, we reference gories. To this end, the company leverages its supe-
additional papers that include more technical details. rior customer insight, which it has gained from its
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
426 Interfaces 46(5), pp. 424444, 2016 INFORMS

Pegboard
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

Shelves
Irregular shelf placement
Table

Polygonal Bins
Chests
Pallets shelf

Figure 1: (Color online) Planograms can have different types of fixtures; as we show in the diagram, examples are
shelves and pegboards. Some planograms have more than one fixture type or present irregular shelf placements,
resulting in irregular planograms.

successful loyalty-card program; this program covers process for a given category is as follows. The cate-
75 percent of Portuguese households and is linked to gory manager (from the commercial department) trig-
approximately 90 percent of Sonae MCs sales. More- gers the process after specifying the product portfolio
over, the company maintains key partnerships with (assortment) for the stores and the key merchandis-
suppliers that have a deep knowledge of its categories ing rules for the portfolios implementation. Product
and assume the role of category captains. Space man- portfolios are not store specific; to manage complex-
agers are also committed to developing planograms ity and effort, they are specified for clusters of stores
with a compelling visual look, and devote a great that have similar sales and space patterns. Space man-
deal of effort to this task. Nevertheless, the attractive- agers start by generating a template planogram (i.e.,
ness of a planogram is subjective and depends on the a role planogram) for each cluster; they use the tem-
space manager who has responsibility for it. Figure 2
plate to determine how merchandising rules fit the
presents an example of a merchandising manual (i.e.,
space. In a collaborative effort, the space and category
a book of rules for planograms to follow, but is orga-
managers then discuss the role planogram and tune
nized using tables) for a category; we can see that
the merchandising rules. When the category manager
products are usually grouped by families, which are
validates the role planogram, then space managers
placed in rectangular shapes. Each planogram has a
replicate it for the remaining stores by adjusting the
hierarchy of families that typically ranges from two
to five criteria. For each criterion, the merchandising product facings to the space of each store, while main-
manual specifies the family type, the display orienta- taining the same allocation rules. Figure 3 summarizes
tion (either vertical or horizontal), the family prece- this shelf-space planning process; it highlights the two
dence, and additional information about preferred key processes: the generation process and the replica-
locations in some cases. tion process. Note that the company has many cate-
The process of updating shelf-space plans involves gories to update and, at the beginning of each year,
ongoing communication between the space managers the category space-planning processes are scheduled
and Sonae MCs commercial department, which is for the entire year.
responsible for managing categories and determin- During the shelf-space updating processes, space
ing their role, depth, and product assortment. The managers generate an average of 60,000 planograms
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
Interfaces 46(5), pp. 424444, 2016 INFORMS 427

Yogurts
Main criteria Classic Yogurts Drinkable Yogurts

BrandHorizontal
2nd TypeVertical
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

Own brand  Economic brand


Criterion Classic  Greek  Low fat
Leader (eye-level)  Sub-leader

BrandVertical FlavorVertical
3rd Leader  Sub-leader  Own brand Strawberry  Mango  Peach
Criterion Economic brand

Precedences Display orientation


Hierarchical Family type (brand, type,
criteria levels flavor, package, size,) (vertical, horizontal)
Preferred locations (eye-level,
hand-level, bottom, top)

Figure 2: Merchandising rules reflect the strategy for implementing a given category; in this figure, the cate-
gory is yogurt. Each manual specifies between two and five hierarchical criteria levels with various types of
information.

Generation process Replication process


Role planogram for each cluster of Planograms for all stores, with the same
stores with same assortment and arrangement as the role planogram
similar sales and space patterns
Store A.1
Assortment A

Role planogram A

Store A.2

Merchandising
rules
Store A.3

Assortment B Store B.1


Role planogram B

Merchandising
rules
Store B.2

Space department with


Commercial department Space department Stores
collaboration of the commercial department

Figure 3: (Color online) The microspace planning process depends heavily on interaction between the space-
planning and commercial departments and comprises two main processes: generation and replication.

each year. To accomplish this, Sonae MC uses space- Thus, it has the necessary capabilities to enable space
planning software from the JDA Software Group, managers to create and maintain the planograms.
Inc. This software includes a space database with all Although automatic tools for planogram genera-
the key information about the products and store tion are available in JDAs software, they do not
equipment, a visualization tool that provides realistic accommodate all the inherent complexity of the mer-
views of the shelves, and powerful reporting tools. chandising rules. Therefore, space managers manually
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
428 Interfaces 46(5), pp. 424444, 2016 INFORMS

developed their planograms by dragging and drop- Theory and Practice of


ping the products onto the shelves, a time-consuming Shelf-Space Management
activity that required an average of three hours for
Most shoppers are susceptible to in-store marketing,
each planogram.
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

mainly because of the low level of involvement that


One of the most difficult challenges that we faced consumers have with in-store decisions. Addition-
at the beginning of the project was the lack of for- ally, reduced assortments and stockouts force them
mal criteria for evaluating planograms. Space man- to search for substitute products, highlighting the
agers were creating and evaluating planograms based role of space management. Experimental studies have
on their intuition and personal judgment, rather than addressed the effect of space variables on the demand
using analytical methods. Nevertheless, in most situ- for products. These studies point to three primary
ations, they were empirically considering space elas- elasticities. (1) Space elasticity measures the increas-
ticities and balancing the product days-supply values. ing responsiveness of demand; as more space is allo-
When analyzing shelf inventory, days-supply is a cated to a product, a decline in marginal returns
common operational metric that measures the num- occurs at some point (Curhan 1972, Chandon et al.
ber of demand days covered by the shelf stock. In 2009). (2) Location elasticity highlights key display
balancing days-supply values, space managers were locations, such as eye level or hand level, that help
using a software highlighting tool that colored the improve exposure (Drze et al. 1994). (3) Cross elas-
products according to predefined days-supply inter- ticity measures the interdependency between adjacent
vals, and they sought to fit all the products within one products and is assumed to be positive for comple-
interval. Moreover, space managers could also have mentary products and negative for substitute products
alternative objectives, such as meeting the brand mar- (Corstjens and Doyle 1981). Additionally, the arrange-
ket shares (i.e., meeting the percentage of total cate- ment of products on the shelves can also be an impor-
gory sales earned by each brand). tant factor in gaining the attention of consumers. Thus,
In 2011, the stores implemented a successful lean carefully organizing products in families can increase
process that, among other modifications, changed interest, while disorganization or excessive complexity
their shelf-replenishment strategy from a just-in-time (i.e., variations in the basic visual content) negatively
policy (i.e., shelves were replenished frequently in affects the buying experience (Pieters et al. 2010).
According to a survey of U.S. retailers (Keltz and
small quantities during the day) to a single shelf-
Sterneckert 2009), the main drivers for space-planning
replenishment operation, which they performed each
initiatives revolve around two main axes: maximizing
day prior to the morning opening. Products normally
selling-space effectiveness, powered by the aforemen-
have joint delivery cycles from the central distribu-
tioned effects, and tighter inventory control. How-
tion centers (i.e., products are delivered on the same
ever, the same survey concludes that the results
days of the week). This fact and the shelf-replenish-
are not meeting the expectations because conven-
ment change of policy explain the reasoning behind
tional assortment analytics and space tools do not
balancing the days-supply values across the various deliver the optimization capabilities needed for suc-
products. By having all products supplied for a simi- cess (p. 1). Software vendors mainly tackle the devel-
lar number of days, the number of shelf-replenishment opment of large-scale data processing technologies
operations decreases, the stock level for long-tail (i.e., capable of addressing shelf-space complexity in prac-
slow-moving) products is controlled more effectively, tice, but with limited or no use of mathematical opti-
stockouts for fast-moving products are prevented, and mization and a complete disregard for the effects of
the backroom inventory (i.e., the number of products consumer demand. Therefore, automatically gener-
in the stores warehouse) can potentially be reduced. ated planograms are still an illusion for most retailers,
Sonae MC believed that analytics could help to who often opt for generic planograms that fit clusters
improve its shelf-space management by going beyond of stores.
a simple planogram automation tool. This provided Shelf-space management, known as the shelf-space
the impetus for our project. allocation problem, is an active field of research
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
Interfaces 46(5), pp. 424444, 2016 INFORMS 429

in retail operations management. Despite the prac- from the backroom and determine individual product-
tical relevance of the problem, scientific knowledge replenishment policies. Our approach can also relate
has been misaligned with practice because most to this stream because we place special emphasis on
state-of-the-art mathematical models have limitations inventory.
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

(Hbner and Kuhn 2012, Bai 2005). Finally, the literature regarding category captains is
The literature presents a great variety of mod- also interesting to this work, and a general treatment
els, most of which differ in their demand functions. can be found in Kurtulus and Toktay (2009). Category
These models incorporate different estimates of some captains are key suppliers that help retailers manage
consumer-demand effects, ranging from complex mul- their categories; for example, they consult on the def-
tiplicative polynomial forms to simplistic linear profit inition of merchandising rules. The use of external
functions. Nevertheless, most of the models have a consultants with in-depth knowledge of the categories
common goal of maximizing demand by determining and the large number of factors behind merchandis-
the product facings on each shelf, without consider- ing rules reinforce the usefulness of rules as input to
ing their locations on the shelves. The most relevant the space-planning process.
approaches to this work are from Corstjens and Doyle
(1981) who were the first to present space elasticity
Automatic Generation of
in a polynomial form, Gajjar and Adil (2010) who
propose a piecewise linearization to the space elastic Planograms (GAP): Overview
demand function, and Yang and Chen (1999) who use This section presents an overview of GAP. More pre-
an alternative model in the form of a linear multiple cisely, we describe the new OR-based process that
knapsack problem. we developed to help space managers at Sonae MC
Perhaps, the most important practical limitation in perform their daily activities more efficiently and
the aforementioned literature is that it neglects mer- effectively.
chandising rules; more specifically, it disregards the GAPs functionality centers on the generation of
existence of product families that specify associations planograms that consider the intrinsic complexity of
between products on the shelves. Russell and Urban the companys shelf-space planning process. In addi-
(2010) and Geismar et al. (2014) are the only authors tion to its analytical advances, one of its most rele-
who define the exact location of the products on vant features is its ability to allow users to control the
the shelves and allocate the space to keep product entire planogram generation process, an activity that
families together in uniform and rectangular shapes. depends highly on the market and is interconnected
Despite this, no model was able to find an optimal with the companys strategy. GAP can incorporate dif-
solution for instances with more than 10 products. ferent types of merchandising rules, which can change
Another key point is that the shelf-space alloca- dynamically to allow space managers to test different
tion literature gives little focus to the cost side of product-layout strategies, and can be tuned to meet
the problem, and most models do not explicitly con- different performance evaluation criteria, such as the
sider inventory-related decisions. Two authors stand equilibrium of days-supply values across the products
out in a more inventory-related stream: Baker and (the most typical planogram evaluation criteria in the
Urban (1988) presented the first model that consid- company) or the maximization of expected demand
ered the demand as a function of the instantaneous based on space-elasticity effects.
inventory level of an item, based on the economic GAPs two major processes, GAP generation and
order quantity model, and Urban (1998) proposed GAP replication, are responsible for the generation
the first attempt to include shelf-space allocation in and replication processes we identify in Figure 3. GAP
the inventory decision-making process. Nevertheless, generation generates planograms from scratch by fol-
these models are comprehensive and are only solved lowing a set of merchandising rules; GAP replication
to optimality for a reduced number of products. replicates a given planogram layout (possibly includ-
The models also include practical limitations: they ing nonstandard manual adjustments) to a new store.
consider continuous shelf-replenishment operations Although the relevance of GAP replication might not
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
430 Interfaces 46(5), pp. 424444, 2016 INFORMS

be clear initially, it was crucial to the success of the optimization models, which we solve using mathe-
project because the large number of fuzzy rules set matical programming-based heuristics (also known as
by the commercial department made the acceptance matheuristics) to ensure fast solutions.
of fully automated planograms unlikely and would Both the GAP generation and GAP replication pro-
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

inevitably lead to manual adjustments. Therefore, cesses start with a thorough data validation and
directly generating store-specific planograms would compatibility analysis to guarantee that all the data
break the validation point in the middle of the pro- are present and are in accordance with the com-
cess. The result would be a need for more validation panys requirements. Once confirmed, the processes
and more manual adjustments by both the category start by calculating the target number of facings for
and space managers. the products, considering one of the possible per-
GAP consists of two main building blocks: (1) GAP formance evaluation criteria. These values are then
optimizer is the systems heart and contains all the used as goals in generating the planograms, either
analytical methods that comprise GAP generation in the planogram generation or planogram replica-
and GAP replication for generating planograms; and tion modules. Finally, because visual attractiveness is
(2) the GAP user interface provides all the tools for important, we designed the last step to improve the
handling data and managing the insertion of mer- planogram aesthetics. If any of the shaded modules
chandising rules. The remainder of the section pro- cannot generate a valid solution, the processes end
vides details on the analytical approach underlying with an infeasibility analysis to determine the causes.
the GAP optimizer, a description of the overall DSS, Next, we describe the key modules, and provide
and an analysis of the key factors that led to a success- a high-level overview of the modeling and solution
ful deployment at Sonae MC. Although GAP consid- approaches included in the GAP optimizer.
ers and integrates many types of fixtures (Figure 1), Target facings. The target facings module deter-
we will focus on the most common type: shelves. mines the number of facings that each product should
have to maximize planogram performance, with-
Analytical Approach out considering any merchandising rules or alloca-
Figure 4 presents the architecture of the GAP opti- tion constraints other than the shelf-space capacity.
mizer. We developed both the GAP generation and The reasons for estimating the target facings before-
the GAP replication processes in a modular fash- hand are twofold: it allows us to consider alternative
ion, and systematically apply innovative tailor-made performance evaluation criteria, and it reduces the

GAP generation

Planogram
generation

Data Target Infeasibility


Aesthetics
validation facings analysis

Compatibility
analysis Planogram
replication

GAP replication

Figure 4: The GAP optimizer has a modular architecture, and some modules are common to GAP generation and
replication. The key modules are in the shaded boxes.
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
Interfaces 46(5), pp. 424444, 2016 INFORMS 431

Days-supply

Shelf level (vertical location)


(a) (b) (c) (d) (e) (f)
interval
Expected
demand
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

Eye-level

Hand-level
Demand

Impact on demand

Shelf space Figure 6: The shelf-attractiveness curves model the attractiveness of the
shelves depending on their vertical locations. We defined six curves (af).
Most of the curves show high attractiveness for the shelves located at eye-
Figure 5: The space-to-sales curve is at the center of the target facings
and hand-levels.
calculation and predicts the demand of a product as a function of its shelf
space, limited to a maximum variation . The vertical lines represent
days-supply intervals.
intervals. For each product, the model determines the
target facings that maximize the planogram expected
complexity of the subsequent allocation problems. For demand, subject to the shelf-space capacity, minimum
brevity, this paper focuses on the companys most fre- and maximum number of facings, and limited to the
quent planogram evaluation objective, which aims to selection of a single days-supply interval. Appendix A
balance the days-supply values across the products, presents additional details about this formulation.
while considering space elastic demand. GAP generation. The GAP generation module deter-
At the center of the target facings calculation is the mines the products placement on the shelves, sub-
space-to-sales curve (Figure 5) that predicts the de- ject to the user-defined merchandising rules. For each
mand of a product as a function of the allocated shelf product, it determines the shelf (or shelves) where
space. This curve reflects the experimental findings the product is to be allocated, its horizontal loca-
regarding the space elasticity effect: the more space tion within the shelves, and the number of facings
allocated to a product, the more consumer awareness to be displayed. Therefore, the outcome is a fully
the product has, leading to increased demand. Nev- defined planogram. This module aims to meet the
ertheless, the marginal returns decrease as the shelf target facings specified upstream (i.e., by the target
space reaches a saturation point, and resembles an S facings module), while considering the location elas-
shape. We include a control parameter in the curve; ticity effects when choosing the products placement.
it specifies the maximum demand variation that can For that purpose, we have defined a set of shelf-
be explained by the shelf space allocated to products, attractiveness curves that model the attractiveness
and considers demand forecasts (given as inputs) as of the shelves depending on their vertical locations
the maximum value. Having captured the space elas- (Figure 6). The shelf attractiveness may vary with the
tic demand, we address the objective of balancing category and fixture type, which explains the alterna-
days-supply values by defining a set of days-supply tive shapes.
intervals, and by limiting all the products to a unique Merchandising rules present a hierarchical structure
interval. The possible days-supply intervals are calcu- of product families, which are a key feature of every
lated in a preprocessing phase using a user-defined shelf-space plan. The products of each family must be
interval length. In Figure 5, we use dashed lines to placed together in adjacent positions; if a family spans
depict these intervals. more than one shelf, products must maintain a con-
The target facings optimization model is formu- tinuous, uniform, and rectangular shape that can be
lated as a mixed-integer program (MIP) and embeds either vertical (similar to columns, which occupy the
piecewise linear approximations of each products full height of the planogram) or horizontal (similar to
space-to-sales curve, obtained using the days-supply lines, which occupy the full length of the planogram).
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
432 Interfaces 46(5), pp. 424444, 2016 INFORMS

Classic Yogurts Drinkable

Level 1
A B
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

Classic Greek Low fat Own brand Economic brand Leader Sub-leader

Level 2 B.4
A.1 A.2 A.3 B.1 B.2 B.3
Sub-leader Economic Sub-leader Economic Sub-leader Economic Strawberry Peach Strawberry Peach Strawberry Peach Strawberry Peach
Leader Own Leader Own Leader Own Mango Mango Mango Mango
Level 3
A.1.1 A.1.2 A.1.3 A.1.4 A.2.1 A.2.2 A.2.3 A.2.4 A.3.1 A.3.2 A.3.3 A.3.4 B.1.1 B.1.2 B.1.3 B.2.1 B.2.2 B.2.3 B.3.1 B.3.2 B.3.3 B.4.1 B.4.2 B.4.3

Level 4
P1 P2 P3

P4 P5

Figure 7: The tree diagram captures all the product-family relationships present in the merchandising rules. This
diagram reflects the merchandising rules for yogurt; see Figure 2.

This hierarchical structure creates complex relation- sequencing variables Tij , stating whether each prod-
ships between the products, which highly constrain uct i precedes or succeeds product j, would increase
the solutions. We capture these relationships using a exponentially as the number of products increases. To
tree diagram (Figure 7). The tree diagram starts with address this, we explore the existence of a hierarchy in
an initial node connecting the families from the first the product families and consider a set of multilevel
criteria that define the first level of the tree. Each fam- family-dependent networks. For each shelf, we start
ily is then connected to its downstream families, lead- by defining a network using the families from the
ing to a multilevel tree of product families. first level of the tree diagram. For each first-level fam-
We formulated this problem using an innovative ily, we then define a network with the corresponding
network-flow MIP model. An intuitive network ap- downstream families and repeat the process until the
proach would associate one node to each product to last level is reached, with a set of disjoint product net-
find the sequence of products on each shelf. How- works. Figure 8 includes a partial definition of the net-
ever, such an approach would lead to an extremely work resulting from the tree diagram in Figure 7. The
complex and intractable model because the traditional model determines the network sequences and then

Network level 1. Shelf k Network level 2 (family A). Shelf k Network level 3 (family A.1). Shelf k

e
A.3 A.1.3 e

A A.2 s
A.1
B A.1.2
A.1.4

e s
s
A.1.1

A A.1 A.3 A.2 A.1.1 A.1.4 A.1.2

Shelf k Shelf k Shelf k

Figure 8: This partial representation of the multilevel family networks results from the tree diagram in Figure 7;
we represent the starting and ending node of each network by s and e, respectively.
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
Interfaces 46(5), pp. 424444, 2016 INFORMS 433

the overall product sequence is obtained by merg- that would consider all family-related merchandising
ing the product-level sequences. This approach also rules and would still be capable of generating high-
guarantees that the products belonging to the same quality, feasible solutions within reasonable time lim-
family are placed consecutively on each shelf, which its. Moreover, using an MIP-based approach enables
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

is another requirement of the problem. Additional us to introduce new features in the problem with lit-
constraints ensure the coordination of each family tle or no effort. We then embedded the formulation in
between shelves for the rectangular shapes and the a matheuristic that can successively solve a sequence
display orientations. The precedence and special loca- of subproblems by exploring the hierarchy present
tion rules, if any, correspond to variables that are fixed in the product families. The matheuristic starts by
during a preprocessing step. allocating families from the first level and progres-
One of the most relevant features of this formu- sively moves down until it reaches the product level.
lation is its flexibility to accommodate any level of Technically, this approach is based on the relax-and-
customization. Nevertheless, in less flexible scenar- fix framework. We consider the entire formulation in
ios that have highly detailed merchandising rules, all iterations; however, families already considered in
the formulation will be more constrained; many vari- previous subproblems have their variables fixed; fam-
ables are fixed and optimization takes place with a ilies not yet considered have their variables relaxed
reduced scope. For example, if the user chooses to set to continuous values; and families from the current
high-level family precedence, the remaining families subproblem have integer variables. A backtracking
are still sequenced by the model. In a more extreme scheme guarantees that whenever a subproblem is
case, if all sequences are defined, the formulation still infeasible, the heuristic shifts backward instead of for-
determines the optimal shelf-space for each product. ward, and solves a larger subproblem by unfixing pre-
As the problem size increases, it becomes intract- vious parts of the solution.
able, even when resorting to the multilevel family Table 1 presents a comparison of the performance
networks. This limits the straightforward use of a of the matheuristic versus the straightforward use
commercial solver on the standard mathematical pro- of a commercial solver on the model formulation
gramming model; that is, it limits using the solver for three real-world generation processes, which we
to solve the entire standard model; other heuris- label CR5 , AI3 , and BH2 ; they correspond to a small,
tic approaches split the model into parts and use medium, and large instance, respectively. Both the
the solver to solve each part, or they remove some solution quality and the run times proved its suit-
constraints. This limitation motivated us to develop ability for use in practice, especially because it gener-
an approximate method. We selected a mathemati- ates high-quality solutions in less than one minute for
cal programming-based approach because it would the small and medium instances and in less than five
be difficult to develop a highly customized heuristic minutes for the larger instance; for the larger instance,

Details Formulation Matheuristic

No. of No. of No. of No. of Optimality


Name products shelves family levels families distance (%) Exec. time Distance (%) Exec. time

CR5 22 5 3 14 0.0 (OS) 00:00:07 0.0 (OS) 00:00:01


AI3 41 9 5 27 8.5 (FS) 01:00:00 001 (FS) 00:00:39
BH2 131 5 4 26 (IS) 01:00:00 (FS) 00:03:54

Table 1: We compared the performance of the matheuristic vs. the straightforward use of a commercial solver
on the standard mathematical programming model for planogram generation (time limit of one hour).

Optimality distance (gap): maximum deviation from the best integer solution to the upper bound; distance:
deviation from the solutions obtained with the matheuristic and a straightforward use of a commercial solver;

(hh:mm:ss); 108% optimality gap, using the models linear relaxation; (OS,FS,IS): (optimal, feasible, infea-
sible) solution.
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
434 Interfaces 46(5), pp. 424444, 2016 INFORMS

the commercial solver failed to deliver a feasible solu- fully complies with the role planogram. In partic-
tion within one hour. ular, it considers the following product placement
In summary, GAP generation is particularly inno- information.
vative for the reasons listed as follows. Products are required to keep the same relative
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

It generates fully defined planograms, which are position as in the role planogram. In the case of shelves,
uncommon in the shelf-space literature. Most mod- this means that products maintain the same shelf level
els address only the space allocation for products and and are placed following the same sequence.
disregard product families, which are an essential fea- Product families are required to keep their uni-
ture in practice. form rectangular shapes. The family continuity within
The mathematical formulation includes model- each shelf is already ensured by keeping the same
ing aspects that significantly improve its performance. sequence. The rectangular shape is obtained by ver-
The most important is the combination of commodity tically aligning the first and the last products of
flow constraints for subtour elimination with the hier- the shape, which we call the left and right align-
archical networks of product families to reduce the ments. This allows us to consider shapes from the role
combinatorial burden of product sequencing. Despite planogram, which may not necessarily be rectangular
being introduced in this context, these modeling fea- (e.g., as a result of manual adjustments).
tures can be explored in other sequencing problems One may say that the role planogram suffers a
that have a hierarchical structure. controlled expansion to keep all alignments. Note
It integrates many practical aspects of merchan-
that during this process, we do not consider location
dising rules not considered previously (e.g., multi-
effects on demand, because the relative product place-
level product families, display orientation, product
ment constrains such a decision.
precedences) and result in more realistic solutions;
Although solving the formulation in a commercial
The suggested matheuristic guarantees the gen-
solver allows us to generate solutions within time lim-
eration of optimized planograms in practical and
its that are acceptable in practice, we developed a
acceptable times even for large instances (e.g., more
second MIP-based heuristic to ensure the scalability
than 200 products).
of the approach, especially for the planograms with
Both the formulation and matheuristic are for-
irregular shelf placements (e.g., from misalignments
mally defined and compared in Bianchi-Aguiar et al.
or interruptions in each shelf level), whose additional
(2015b). Appendix B includes an overview of the
mathematical formulation. constraints greatly impact the performance of the
GAP replication. Given a fully defined planogram formulation.
(i.e., role planogram), the GAP replication module Generically speaking, this matheuristic has three
reproduces a similar product placement for a new main steps. The first step generates an initial solution
store without the need to provide merchandising for the problem with the minimum display quanti-
rules or other types of reasoning behind the plano- ties for the products. The second step iteratively adds
gram construction. The new space is usually larger the remaining product facings to the planogram until
in width (planograms usually have the same height); no more space is available (or no more facings can
however, it should have a similar shelf layout to be added). The last step tries to improve the solu-
ensure compatibility between the two planograms. tion by allowing the removal and insertion of new
Because the new store would typically have a differ- facings. Technically, this approach is an integration of
ent demand pattern, the objective is to meet a new tar- two well-known MIP-based improvement heuristics:
get number of facings, which are specified upstream fix and optimize and local branching. Thus, in each
and are suitable for this store. iteration, we solve the model with some variables par-
We formulated the replication problem as an MIP tially constrained in one of two ways: a subset of the
model. Although we aim mainly to adjust the product variables is fixed to the values obtained in the incum-
facings, the model necessarily determines the prod- bent solution (i.e., fix and optimize) or a limited num-
ucts locations to guarantee that the new planogram ber of changes is allowed to the values obtained in the
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
Interfaces 46(5), pp. 424444, 2016 INFORMS 435

Details Formulation Matheuristic

No. of No. of No. of Optimality


Name products left alignments right alignments distance (%) Exec. time Distance (%) Exec. time
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

E1 49 3 4 0.0 (OS) 00:00:14 001 (FS) 00:00:03


J1 107 11 10 0.0 (OS) 00:01:27 002 (FS) 00:00:02
G1 240 32 35 0.12 (FS) 01:05:00 002 (FS) 00:00:24

Table 2: We compared the performance of the matheuristic vs. the straightforward use of a commercial solver
on the standard mathematical programming model for planogram replication (time limit of five minutes).

Optimality distance (Gap): maximum deviation from the best integer solution to the upper bound; Distance:
deviation from the solutions obtained with the matheuristic and a straightforward use of a commercial solver
(OS,FS,IS): (optimal, feasible, infeasible) solution.

incumbent solution (i.e., local branching). One inter- the current paper, we provide an overview of the
esting aspect of this matheuristic is that it mimics the mathematical formulation in Appendix C.
process that the space managers follow when they Aesthetics. The formulations for planogram genera-
manually replicate planograms. tion and replication focus on ensuring that the shapes
Table 2 compares the results obtained using our are rectangular and disregard other aesthetic details.
solution approach and a commercial solver on the As a result, some planograms have display issues,
model formulation. We again use three real-world such as large and irregular gaps between the prod-
instances ranging from small to large (E1 , J1 , and G1 ), ucts. The aesthetics module is responsible for improv-
to illustrate the efficiency and scalability of the ing the attractiveness of the planogram. It considers
approach. two key factors to obtain attractive displays: (1) the
In summary, GAP replication is particularly inno- spacing of the products throughout the planogram,
vative for the following reasons. and (2) whether the planogram is fully merchan-
We introduce the replication problem in the shelf- dised (i.e., full of facings). Therefore, it executes the
space literature, presenting a novel, practice-oriented generation or replication formulation, depending on
mathematical formulation. Most retailers who are whether the process is GAP generation or replication,
forced to cluster their stores to efficiently manage respectively; all the decisions are fixed to the incum-
categories use the replication process. Therefore, we bent solution, with the exception of the horizontal
believe that this is an important step toward the use location of the products. This module changes the
of analytic methods in shelf-space practice. objective function to (1) minimize the empty space,
For retailers with complex merchandising rules, and (2) minimize the maximum spacing between two
this is a valuable tool that allows them to translate consecutive products when no more products fit the
these hard-to-grasp and fuzzy rules into one example planogram. This latter objective distributes the empty
that they can then use as the basis for creating the space throughout the products.
remaining planograms. Infeasibility analysis. Highly customized and de-
The three-phase mathematical programming- tailed merchandising rules lead to significantly con-
based heuristic can solve most problems in less than strained generation and replication formulations,
three minutes, thus guaranteeing the robustness and which can compromise the existence of a feasible
scalability of our approach. It does this by combin- solution for the problems. Moreover, the target facings
ing state-of-the-art programming heuristics, such as formulation can also be infeasible, thus resulting in
fix and optimize and local branching, in a single too many possible reasons for the process to end with-
framework. out a valid solution. To overcome these data-related
Both the formulation and matheuristic are for- issues, we developed an infeasibility analysis mod-
mally defined and compared in Bianchi-Aguiar et al. ule that searches for the possible infeasibility causes
(2015a). To ensure the readers the comprehension of using a structured and logical procedure. It performs
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
436 Interfaces 46(5), pp. 424444, 2016 INFORMS

multiple runs of the infeasible formulation; for each GAP user interface, which manages the data-handling
run, it removes a problem feature or requirement from process, executes the replication and generation pro-
the formulation. The process stops after identifying a cesses with real-time status messages, and presents
source for the infeasibility (i.e., whenever the formu- the generated planograms and a full execution report.
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

lation is able to find a valid solution for the relaxed This report provides all the warnings and errors that
problem). occurred during the process and, when applicable, the
infeasibility causes. That is, this interface is present
Decision Support System throughout the entire process to help with interact-
The GAP optimizer requires the integration of various ing with the users. Figure 9 depicts the most relevant
types of information obtained from multiple sources. flows of information and shows snapshots from two
If this information is not handled carefully, it may interface forms: the project manager for handling the
jeopardize the successful use of the application. To data and the generation manual for managing rules
this end, another important building block is the and configurations.

Product Solution
assortment report

Users
Equipment

Planogram(s)
Product
performance Project manager
IKB database (JDA)
IKB database (JDA)

Merchandising
rules

Configurations

Users

Generation/Replication manual

Inputs GAP user interface Output

GAP Commercial
optimizer solver

Figure 9: (Color online) GAPs two main building blocks, the GAP optimizer and GAP user interface, require the
integration of information obtained from multiple sources.
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
Interfaces 46(5), pp. 424444, 2016 INFORMS 437

The generation manual is one of the systems key We strongly believe that some factors associated
components. Inspired by merchandising manuals, as with GAPs design and project management were cru-
the example in Figure 2 shows, this form presents a cial to the projects success. From the perspective of
familiar interface (to space managers) so that they can both the space and commercial departments, the deci-
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

configure rules. Space managers have a high degree sion to divide GAP in two processes was critical,
of flexibility in defining these rules; in each run, they because GAPs development could not disrupt cur-
can choose the level of customization that they require rent practices. Starting the implementation with GAP
in the generated planograms. For the advanced users, replication was also in hindsight a wise decision for
several other configurations are available; these can two reasons. First, the replication process was quicker
range from alternative-location elasticity curves and to implement and provided more consensual solu-
planogram evaluation criteria to control parameters tions; therefore, it helped to bring the space man-
and tolerances. Each generation manual can be saved, agers on board with GAP at an earlier stage. Second,
consulted, and reused in multiple processes. Most it allowed us to deepen our knowledge of the com-
importantly, it can evolve as the space managers eval- plex structure in which the merchandising rules are
uate planogram solutions and make changes to the defined.
planograms. The close collaboration between the three space
The GAP Optimizer is a C++ program in which managers on the project team was also essential,
all the models are embedded in the code. Sonae MC because they provided valuable contributions through-
acquired a commercial solver, and the formulations out the course of the project, for example, during
are executed using a C++ library from this solver. requirements gathering, problem definition, testing,
We developed the GAP user interface using Win- and validation. Weekly meetings between FEUP and
dows forms, and all communications between the the space managers were important for maintain-
two building blocks use XML files. Both the GAP ing open communications on project-related issues
optimizer and the GAP user interface are executed and ensuring that project milestones were met. This
on a dedicated server and all space managers have continuous process allowed us great flexibility in
access to the interface using a remote desktop con- implementing GAP and enabled us to develop an
nection in a terminal-server architecture. Currently, application tailored to Sonae MCs requirements.
GAP does not have a direct connection to the space
Additionally, the three space managers tested and val-
database, and the information is manually exported
idated GAP using different categories of products;
and imported to the interface. Given the success of
this was a significant factor in building their internal
the project, Sonae MC is now studying more effi-
confidence in the application and communicating it
cient infrastructures, both for communicating with the
within the company.
space-planning database and for accessing the server
Nevertheless, we faced some challenges in this col-
from the space managers terminals.
laboration between academic researchers and indus-
Project Development try practitioners, because they have different objec-
The project was launched in March 2012 and ran until tives, incentives, and time horizons. At the beginning
July 2014. The two processes, GAP generation and of the project, we found that the space managers were
GAP replication, were developed sequentially, and highly resistant to adopting the new process; how-
each involved three main phases: requirements def- ever, we were able to gradually overcome their resis-
inition, prototype and proof of concept, and testing tance by attempting to keep them updated on the evo-
and validation. From an organizational standpoint, it lution of the project. This improved their commitment
included a team from FEUP, responsible for the devel- to GAP and allowed them to understand the applica-
opment of GAP, and two Sonae MC teams: a team tions potential.
from its space department, responsible for validating We also gave training sessions before the rollout of
requirements and testing GAP, and a team from its ISI each of the two processes; given the systems com-
department, responsible for integrating GAP into the plexity, this training was crucial to engaging space
companys information systems. managers. These sessions included the analysis of
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
438 Interfaces 46(5), pp. 424444, 2016 INFORMS

planograms with unexpected characteristics obtained after they become more familiar with the new soft-
from using GAP and the development of a check- ware and totally, or partially, reuse the generation
list for systematically looking for alternative solutions manuals that they carefully developed for the first
in these situations. Moreover, the execution report processes.
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

also played a vital role in allaying the concerns Additionally, GAP shifted the space managers
of space managers when GAP generated an unex- focus from planogram construction to planogram
pected solution and, more importantly, when it did evaluation, allowing them to concentrate on addi-
not generate one. tional activities, such as market-trend studies and
Last, and perhaps one of the most critical factors experiments with alternative merchandising rules.
for the success of the GAP implementation, the space Therefore, this change of paradigm brought increased
planning and innovation directors showed enormous responsibility to space managers, and gave them an
commitment and provided their full sponsorship dur- analytical tool to support their decisions during meet-
ing the entire project. ings with the commercial department.

Impact Optimization: Targeting Optimality for


All Customization Levels
Today, the entire micro-space team at Sonae MC uses
The impact of GAP generation and GAP replication
GAP daily for both for the generation and repli-
on optimization must be measured independently.
cation of planograms. In this section, we describe
One of the greatest advantages of GAP generation
how it enhanced shelf-space management for each
is its flexibility to generate either highly customized
of the three axes that we identified for the project:
solutions or more demand-driven (and concurrently
process automation, space optimization, and display
innovative) solutions, based on the specifications in
standardization.
the generation manual. Figure 10 depicts this flexi-
Automation: From Planogram bility by showing three planograms: the first was
Construction to Evaluation generated manually by a space manager and the re-
Perhaps, the most important impact of the project maining two were generated with GAP, first using a
was its effect on the space-management process: it fully defined generation manual (high customization)
improved processing times and resulted in a positive and then using the same manual but removing family
change of paradigm. precedences and display directions (low customiza-
During the first months after the rollout, all space tion). The first generated planogram is largely similar
managers were encouraged to use GAP in their daily to the manually generated planogram; the second
tasks, and to register the number of planograms that presents the alternative reasoning behind its creation,
GAP could generate and the process execution times, where the intention was to put the most popular fam-
including the duration of data handling, creation of ilies in the premium vertical and horizontal locations.
the generation and replication manuals, and man- Regardless of the customization level, GAP gener-
ual adjustments made to the final solution. When we ation always uses the available degrees of freedom
compared the execution times with the legacy process, to optimize the number of product facings and the
we found that the results were encouraging. Based products locations. To assess the impact of GAP gen-
on an analysis of over 400 planograms in eight cat- eration on the planograms performance, we analyzed
egories, space managers were able to automatically 25 generation processes (one of which was the exam-
generate 80 percent of the planograms and the cat- ple from Figure 10). These examples were carefully
egory space-management processes took on average selected during the proof-of-concept phase to guaran-
48 percent less time (46 percent less in the generation tee that all specificities of the categories were covered.
processes and 50 percent less in the replication pro- We evaluated the impact by measuring the follow-
cesses). Moreover, space managers highlighted that ing performance metrics: potential sales increase (esti-
these time reductions might improve in the future, mated using the location elasticity curves with a
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
Interfaces 46(5), pp. 424444, 2016 INFORMS 439

(AND MADE
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

PLANOGRAM
6ERTICALIMPACT

'!0PLANOGRAM
HIGHCUSTOMIZATION
6ERTICALIMPACT

'!0PLANOGRAM
LOWCUSTOMIZATION

(ORIZONTALIMPACT (ORIZONTALIMPACT
(IGHLIGHTLEVEL (IGHLIGHTLEVEL

Figure 10: (Color online) In this figure, we show two planograms generated by GAP using two levels of cus-
tomization, and we compare them to a manually generated planogram.

maximum impact of 20 percent); days-supply bal- GAP generation


ance (measured in terms of the average and the stan- Low High GAP
dard deviation reductions); and planogram filling rate customization customization replication
(defined by the ratio of the linear space utilized and
Potential 104 007
the overall available space). Table 3 summarizes the
sales increase (%)
results when generating these planograms using high Average days-supply 3706 3503 3403
and low customization levels. The percentage values reduction (%)
Standard deviation 6008 5104 5603
are relative to the manually generated planogram. days-supply reduction (%)
GAP generation improves the manual planogram per- Space occupation (%) 9605 9607 9703
Execution time (hh:mm:ss) 00:07:12 00:01:50 00:02:10
formance in all of these metrics, both in the low
and high customization versions. As expected, reduc-
Table 3: To assess the impact of GAP, we analyzed 25 generation and
ing the number of rules imposed on the planogram replication processes. Generation processes were run with low and high
yields additional gains with an average increase of customization settings. Both processes bring benefits to the current prac-
tice, both in terms of potential sales increase and inventory management.
potential sales from 0.4 to 1.4 percent. We observed
With respect to the manually generated planogram; 94% in manual
that the gains were more relevant in the metrics planogram.
regarding the day-supply values (whose average and
standard deviation decreased by 38 percent and 61 planograms filling rate also increases by three percent
percent, respectively), which is consistent with our compared to the manually generated planograms.
primary objective. Note that days-supply values have GAP replication, by definition, has less margin for
a major impact on replenishment operations, holding optimization, because changes are allowed to only
costs, and product availability. In both versions, the the number of facings that products have in the new
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
440 Interfaces 46(5), pp. 424444, 2016 INFORMS

planograms, subject to the many allocation rules that its complexity and high dependency on merchandis-
were extracted from the role planogram. Neverthe- ing rules. We believe that this work is an important
less, a better product-facing allocation can optimize contribution in this area from both theoretical and
the days-supply values. We used the same 25 exam- practical points of view. On the scientific front, we
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

ples to assess GAP replication performance by repli- provide innovative mathematical models and efficient
cating the handmade planograms to the same store. algorithms for the shelf-space allocation problem and
The results in Table 3 prove that we are still able to bring more practicality to the scientific approaches to
significantly improve days-supply balancing. this problem. From the application perspective, we
This project also resulted in a significant cultural give insights on how to tailor analytical approaches to
change within the company. The success of the project the practice of shelf-space management. We introduce
motivated the use of OR approaches (and more gener- the replication problem and allow users to control
ically speaking, of analytical approaches) in other the level of customization from solutions, while still
operations planning activities. In particular, it trig-
applying optimization at every step of the process.
gered many other projects with the same OR group
For Sonae MC, this resulted in automated processes,
from the University of Porto, both in space-related
which were able to take advantage of the space man-
problems, such as backroom optimization, and in
agers experiences and knowledge. We also provide
related areas, such as marketing, store operations, and
project management details that were critical during
logistics.
GAP implementation at Sonae MC, the Portuguese
Standardization: Knowledge Management for a retail company that partnered with us in this project.
Global Process Although this paper describes a real application of
Space managers are divided into groups responsible shelf-space planning, the approach does not intrin-
for subsets of categories. The knowledge associated sically depend on any company-specific policies,
with each category is within each group, and is sup- because it is based on rules that are defined at run
ported by manuals that record the implementation time. Therefore, it is sufficiently generic to be suitable
details (using a template similar to the template in for other retail companies in the grocery industry or
the example in Figure 2). Nevertheless, these manu- in similar industries. Its modular nature also enables
als are frequently limited to the upper-criteria levels, its adaptation and integration with other information
and give only a general idea of the reasoning behind technology systems.
the planograms. Consequently, most of the category
knowledge resides with the space managers, and can
be partially lost when organizational changes occur.
Acknowledgments
GAP also had a major impact on standardizing infor- The authors thank the other members of the team. They give
mation and managing knowledge. First, the use of special thanks to the three space managers, Constantino
electronic generation manuals (and the possibility of Gomes, Pedro Soares, and Susana Borges, and to Frederico
reusing them) centralized the space-planning knowl- Santos, Joel Pacheco, Miguel Camanho, and Hlder Matos
edge associated with each category, and made it pos- from ISI. The authors also acknowledge the relevant com-
sible to systematize the tacit knowledge available into ments from the 2015 Daniel Wagner Prize judges, with spe-
cial thanks to Ann Bixby, and to J.H. Tice for the English
information that could be shared among peers. Sec-
revision.
ond, it reduced the subjectivity of the process, which The authors also thank the Fundao para a Cincia e a
is now less dependent on the managers experiences. Tecnologia (FCT, the Portuguese Foundation for Science and
Technology) for awarding grants [SFRH/BD/74387/2010
Concluding Remarks and SFRH/BPD/98981/2013]. This work is also financed
by the North Portugal Regional Operational Programme
In todays highly competitive retail environment, re-
(NORTE 2020), under the PORTUGAL 2020 Partnership
tailers can benefit from analytic tools that improve Agreement, and through the European Regional Develop-
decision making; the literature reports on many suc- ment Fund (ERDF) [Project TEC4 GrowthPervasive Intel-
cessful examples. Shelf-space planning is one area ligence, Enhancers and Proofs of Concept with Industrial
that still has much to be explored, mainly because of Impact/NORTE-01-0145-FEDER-000020].
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
Interfaces 46(5), pp. 424444, 2016 INFORMS 441

Appendix A. Target Facings Model dsin n Win dsin+1 n+1 1 i N1 n T1


In this appendix, we provide a mathematical formulation (number of facings)3 (A4)
of the target facings model. Consider a specific category X
of a store with overall capacity C. The retailer wants to n = 11 (single days-supply interval)3 (A5)
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

nT
allocate N products, indexed by i N, with length ai . Each
product is associated with a space-to-sales curve (Figure 5), Win 0 1 i N1 n T3 n 801 191 n T1
which is linearized with piecewise lines that are obtained (integrality)0 (A6)
using the facings associated with the days-supply intervals
(Figure A.1). There are T days-supply intervals, indexed by
n T. For each product i, the minimum and maximum fac- Appendix B. Generation Model
ings of each interval n are dsin and dsin+1 . In this appendix, we provide a simplified mathematical for-
This space-to-sales curve (Figure 5) is used widely in mulation of the generation model. The notation used in
the literature and is associated with a polynomial func- Appendix A is extended by introducing K shelves, indexed
tion that depends on the space-elasticity parameter that by k K and with length wk . The shelf attractiveness of
Corstjens and Doyle (1981) first introduced. Other authors, Figure 6 is modeled by parameter k . Additionally, pi is the
such as Gajjar and Adil (2010), have proposed piecewise lin- profit per facing of product i and li and ui are the lower
ear approximations. However, we are the first to consider and upper bounds, respectively, for the number of facings
the problem with days-supply intervals. of product i. These bounds are determined by the upstream
The objective is to maximize the planograms expected target facings model and correspond to the bounds of
demand by determining the number of facings for each the selected days-supply interval (li = n n dsin and
P
product i. The decisions to be made are: n , which speci-
P n+1
ui = n n dsi ).
fies whether the days-supply interval n is selected for all The objective is to maximize the planograms expected
products; and Win , which indicates the number of facings of profit by determining the number of facings for each prod-
product i if the days-supply interval is n . uct i and its placement on the shelves, while respecting all
The formulation is as follows: merchandising rules, translated into a tree diagram. Let Wik
XX n n be the integer number of facings of product i on shelf k. The
Maximize fi 4Wi 5 (A1) objective function and product-allocation constraints can be
iN nT
defined as follows:
subject to: XX
Maximize Z = pi k Wik (B1)
Win ai C1
XX
(shelf-space capacity)3 (A2) iN kK
iN nT
subject to:
Win ui 1
X
li i N1 X
nT li Wik ui 1 i N1 (minimum and maximum
(minimum and maximum facings)3 (A3) kK
number of facings)3 (B2)
X
ai Wik wk 1 k K1 (shelf capacity)0 (B3)
%XPECTED iN
DEMAND F 4I 7I
FIN  7I To ensure a correct positioning of the products within
F IN 7I its product family and on the shelf, let u M denote the

set of product families and Nu the set of products of fam-
$EMAND

F I 7I ily u. To capture the tree diagram of Figure 7, consider that


 the set Vu contains the immediately downstream families
F I 7I FI 7I
of parent family u, which for the final level corresponds
to products. The family sequencing in the planogram is
ensured by the multilevel network-flow constraints pre-
sented next, where Ymk (0/1) defines if family m (or prod-
DSI DSI DSI DSIN DSIN  DSIN  uct i) is located on shelf k and Tmnk (0/1) captures if
family m is displayed immediately after family n on shelf k.
   N N  4
X
3HELFSPACEOFPRODUCTI 7I T0mk = Yuk 1 u M1 k K1 (first family belonging
mVu
to parent family u displayed in the shelf)3 (B4)
Figure A.1: (Color online) The space elasticity curve is linearized using X
piecewise linearization based on the days-supply intervals represented by
Tm0k = Yuk 1 u M1 k K1 (last family belonging
mVu
the vertical lines. to parent family u displayed in the shelf)3 (B5)
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
442 Interfaces 46(5), pp. 424444, 2016 INFORMS

Xms Xus +
X X
Tnmk = Ymk 1 u M1 m Vu 1 k K1 Fnmk + C 41 Ymk 51
nVu 809 nVu 8092 n6=m
(only one family immediately before)3 (B6)
u M1 m Vu 1 k K1 (left alignment)3 (B17)
X
Tmnk = Ymk 1 u M1 m Vu 1 k K1
Xme Xms
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

X
nVu 809
Lik 1 u M1 m Vu 1 k K1
(only one family immediately after)3 (B7) iNm
(right alignment)3 (B18)
Ymk Yuk 1 u M1 m Vu 1 k K1 (only on the shelf if
Xme Xms v
X
Lik + C 41 Ymk 51
the parent family is allocated)3 (B8) iNm

Wik Yik 1 i N1 k K1 (only display facings on the u M1 m Vu 1 k K1 (right alignment)3 (B19)


shelf if the product is allocated)0 (B9) X
F Lmk = 11 F Lm0 = Ym0 1 u M1 m Vu 1
kK
To complete the sequencing decisions and distribute the (first-shelf identification)3 (B20)
shelf length among the several product family layers as X
depicted in Figure 8, the following single-commodity con- LLmk = 11 LLmK = YmK 1 u M1 m Vu 1
kK
straints are required: (last-shelf identification)3 (B21)
X
F0mk = 01 u M1 k K1 F Lm1 k+1 + Ymk = Ym1 k+1 + LLmk 1 u M1 m Vu 1
mVu
(null initial flow in each network)3 (B10) k K2 k 6= K1 (shelf consecutiveness)0 (B22)
X X
Fm0k = Lik 1 u M1 k K1 Bianchi-Aguiar et al. (2015b) provides additional details
mVu iNu about the generation model.
(the end flow is equal to the sum of the
family products length)3 (B11)
Appendix C. Replication Model
In this appendix, we provide a simplified mathematical for-
X X X
Fnmk + Lik = Fmnk 1 u M1
nVu 8092 m6=n iNm nVu 8092 m6=n mulation of the replication model. We will follow the nota-
tion used in Appendices A and B.
m Vu 1 k K1 (flow-conservation constraints)3 (B12)
The objective is to replicate the product placement of
Fmnk wk Tmnk 1 u M1 m1 n Vu 8092 m 6= n1 k K1 a given planogram to a new store without the need to
specify merchandising rules, while maximizing the new
(flow can only traverse the selected arcs)3 (B13)
planograms expected profit. For that, products should
X
Lik = wk 1 k K1 be assigned by the same sequence to the same shelves
iN and product families should keep their rectangular shapes,
(all shelf lengths must be assigned)3 (B14)
which can be ensured by aligning the familys first and last
ai Wik Lik 1 i N1 k K1 (the shelf length assigned product of each shelf (with a tolerance of v units). To accom-
plish that, consider the sets Nk containing the products from
must hold the facings to be displayed)3 (B15)
each shelf k, M L left alignments, and M R right alignments,
where Lik represents the shelf length assigned to product i indexed by m MR ML . Because products cannot change
on shelf k and Fmnk is the cumulative assigned shelf length between shelves, the shelf attractiveness was incorporated
(commodity flow) when changing from family m to fam- in the profit: pi
ily n on shelf k. This is used to describe the networks Let the products be numbered according to their position-
of Figure 8. Note that constraints (B14) and (B15) replace ing. Wi is the integer number of facings of the ith product
constraints (B3). and Li is the shelf length assigned to the ith product i. Addi-
The final set of constraints guarantee the rectangular tionally, XmL and XmR capture the horizontal location of left
shape of product families in the planogram. We impose the and right alignments, respectively. The replication model
rectangular shapes by determining the left and right coordi- reads:
nates of each family and ensuring that the familys first and X
last products of each shelf are kept within v units from these Maximize Z = pi Wi (C1)
iN
coordinates. To complete the rectangular shape, we force
the family to be in consecutive shelves. To do so, let Xms subject to:
(Xme ) be the left (right) horizontal coordinate of the family m
and F Lmk (LLmk ) be equal to 1 if k is the first (last) shelf of li Wi ui 1 i N1 (minimum and maximum
block m. number of facings)3 (C2)

Xms Xus +
X
Fnmk 1 u M1 m Vu 1 k K1 Li ai Wi 01 i N1 (the shelf length assigned
nVu 8092 n6=m
(left alignment)3 (B16) must hold the facings to be displayed)3 (C3)
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
Interfaces 46(5), pp. 424444, 2016 INFORMS 443

X
Li = wk 1 k K1 (all shelf lengths must Kurtulus M, Toktay LB (2009) Category captainship practices in
iNk be assigned)3 (C4) the retail industry. Agrawal N, Smith SA, eds. Retail Supply
Chain Management: Quantitative Models and Empirical Studies
(Springer, New York) 7998.
X
Xi = Lj 1 k K1 i Nk 1 (product positioning
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

jNk 2 j<i Pieters R, Wedel M, Batra R (2010) The stopping power of adver-
and relationship)3 (C5)
tising: Measures and effects of visual complexity. J. Marketing
74(5):4860.
XmL v Xis 1 m ML 1 i NLm 1
Russell RA, Urban TL (2010) The location and allocation of prod-
(product families left alignment)3 (C6) ucts and product families on retail shelves. Ann. Oper. Res.
179(1):131147.
XmL Xis 1 m ML 1 i NLm 1 Urban TL (1998) An inventory-theoretic approach to product assort-
(product families left alignment)3 (C7) ment and shelf-space allocation. J. Retailing 74(1):1535.
Yang MH, Chen WC (1999) A study on shelf space allocation and
XmR + v Xis + Li 1 m MR 1 i NRk 1 management. Internat. J. Production Econom. 61(510):309317.
(product families right alignment)3 (C8)
Teresa Bianchi-Aguiar is senior consultant at LTPlabs,
XmR Xis + Li 1 m MR 1 i NRk 1
a management consulting company that applies advanced
(product families right alignment)0 (C9) analytical methods, and a researcher at the Center for
This formulation requires each product to stay in a single Industrial Engineering and Management of INESC TEC.
product. If this is not the case, then each products position She holds a PhD in industrial engineering and manage-
should be assigned a different index and small changes are ment from the University of Porto. Her main interests
needed in the formulation. Bianchi-Aguiar et al. (2015a) pro- are applying analytics into complex real-world problems
vides additional details about the replication model. from companies, specifically within the retail industry. She
has experience in shelf-space planning, demand forecasting,
References and inventory management.
Bai R (2005) An investigation of novel approaches for optimising Elsa Silva is researcher at the Centre for Industrial Engi-
retail shelf space allocation. Doctoral dissertation, The Univer- neering and Management of INESC TECTechnology and
sity of Nottingham, Nottingham, UK. Science Associate Laboratory. Her main research interests
Baker R, Urban TL (1988) A deterministic inventory system with include the ability to solve hard and large combinatorial
an inventory-level-dependent demand rate. J. Oper. Res. Soc.
39(9):823831. optimization problems by using hybrid approaches of linear
Bianchi-Aguiar T, Carravilla MA, Oliveira JF (2015a) Replicating programming and metaheuristics. The main area of appli-
shelf space allocation solutions across retail stores. Euro. J. Oper. cation of her research is cutting and packing problems.
Res. Submitted. Luis Guimares is a professor in the Department of In-
Bianchi-Aguiar T, Silva E, Guimares LG, Carravilla MA, Oliveira
dustrial Engineering and Management of the Faculty of
JF (2015b) Allocating products on shelves under merchandis-
ing rules: Multilevel product families with display directions. Engineering at the University of Porto; researcher in the
Omega. Submitted. Industrial Engineering and Management group at INESC
Chandon P, Hutchinson JW, Bradlow ET, Young SH (2009) Does in- TEC; and co-founder of LTPlabs. His main area of activity is
store marketing work? Effects of the number and position of operations research/computer science. Most of his research
shelf facings on brand attention. J. Marketing 73(6):117.
Corstjens M, Doyle P (1981) A model for optimizing retail space
is problem driven and aims to develop advanced analytic
allocations. Management Sci. 27(7):822833. solutions to be applied to real-world problems. He has col-
Curhan RC (1972) The relationship between shelf space and unit laborated in several industry-based research and consult-
sales in supermarkets. J. Marketing Res. 9(4):406412. ing projects with various companies in the areas of process
Drze X, Hoch SJ, Purk ME (1994) Shelf management and space
industry, transportation, retail, and energy.
elasticity. J. Retailing 70(4):301326.
Gajjar H, Adil G (2010) A piecewise linearization for retail shelf Maria Antnia Carravilla is professor of OR/MS at the
space allocation problem and a local search heuristic. Ann. University of Porto, Faculty of Engineering, visiting pro-
Oper. Res. 179(1):149167. fessor at Universidade de So Paulo (USP), and a senior
Geismar HN, Dawande M, Murthi B, Sriskandarajah C (2014) Max- researcher at INESC TEC. Her main area of research activ-
imizing revenue through two-dimensional shelf-space alloca-
ity is OR/MS. She develops analytical models and methods
tion. Production Oper. Management 24(7):11481163.
Hbner AH, Kuhn H (2012) Retail category management: State- for combinatorial optimization problems arising in various
of-the-art review of quantitative research and software appli- domains (retail, mobility, healthcare, and manufacturing)
cations in assortment and shelf space management. Omega and was responsible for several successful contracts with
40(2):199209. industry and services, and R&D projects addressing these
Keltz H, Sterneckert K (2009) The trend toward consumer
Centric merchandising requires assortment management and problems.
space planning investments. Technical report, AMR Research, Jos F. Oliveira is full professor in the Department of
Boston. Industrial Engineering and Management of the Faculty of
Bianchi-Aguiar et al.: Analytics to Enhance a Food Retailers Shelf-Space Management
444 Interfaces 46(5), pp. 424444, 2016 INFORMS

Engineering at the University of Porto and collaborates with in different business sectors including manufacturing, ser-
the Business School of the University of Porto. He is senior vices, and retail, and is part of the Sonae group since 2001.
researcher at the Industrial Engineering and Management Jorge Liz is the responsible for the Space Planning team
group at INESC TEC. His primary research interests are of Sonae MC. Working at Sonae since 2000, he started by
Downloaded from informs.org by [115.113.11.143] on 29 January 2017, at 09:22 . For personal use only, all rights reserved.

decision and optimization problems related to the efficient leading the planning and control department, and then
use of raw materials and other resources (cutting and pack- started a zig-zag career at Sonae that included positions at
ing problems) and decision support systems in industry and almost all the Sonae businesses (food and nonfood retail,
services. He is editor of the European Journal of Operational wood based panels, private equity, tourism, insurance, real
Research. estate).
Joo Gnther Amaral is a member of the Executive Srgio Lapela is responsible for the Micro Space and the
Board of Sonae MC with the responsibility of Logistics System and Processes Integration Division at Space Depart-
and Production Units and Supply Chain Management Ini- ment. After 20 years at Sonae MC, he now has a team of
tiatives. He was previously the Head of Innovation at 23 FTEs, responsible for space management, development
Sonae, leading the implementation of several innovation and innovation on food, nonfood, and fresh departments.
projects in tight articulation with the different business He was also involved in various internal projects in part-
units. He has more than 20 years of IT experience working nership with FEUP, Fraunhofer Portugal.

You might also like