DΛTΛ FOR

RETΛIL
 WE UNLEASH THE FULL VALUE OF
DATA THROUGH DEMOCRATIZATION

The Netherlands

UK Germany
France Switzerland
Spain
New York South Korea
Lebanon
Morocco
Saudi Arabia China
UAE
Los Angeles Mexico

Colombia Malaysia
Singapore

Brazil

18 +1300 +1000
COUNTRIES EMPLOYEES CLIENTS

Artefact is a global leader in consulting services, specialized in data transformation

and data & digital marketing, from strategy to the deployment of AI solutions.
We are offering a unique combination of innovation (Art) and data science (Fact).

DATA ACCELERATION PROGRAMS | AI SOLUTIONS | DATA MARKETING

 DATA FOR RETAIL

TABLE OF CONTENTS

Data for Retail

4 Introduction by Vincent Luciani CEO & co-founder of Artefact

7  ecoming data-driven (and reaping

B 36 MATTEL × RelevanC Advertising —
immediate benefits) has never been How Artefact helped boost Mattel’s
easier in the retail sector online sales on Cdiscount (RelevanC
Advertising) retail media platform?
9 CARREFOUR GROUP — How Data & AI
can accelerate sustainable business 38 CARREFOUR GROUP — How to reduce
transformation food waste in the Bakery-Pastry
department?
12 
Data-driven marketing: the rise of the
customer data platform 41 
Demand forecasting: Using machine
learning to predict retail sales
15 CONFORAMA AI-enabled
personalization boosts Conforama 43 
Sales forecasting in retail: what we
CRM campaign revenues learned from the M5 competition

19  etail Media: An indispensable asset

R
for brands

21 UNILEVER — How does Artefact

support UNILEVER on Retail Media
use cases to increase its sales?

31  ata monetization opportunities for

D
retailers: Retail Media within the CPG/
Retailer data ecosystem

34 CARREFOUR LINKS — With Artefact

are helping brands to improve media
targeting thanks to Carrefour’s data

3
Interview with Vincent Luciani about
Data & AI market perspectives
The strategic importance of data for companies is no
longer in question. Aware of this reality, Artefact helps
companies to capitalize on this performance, growth
and improvement lever.

How would you analyze the

latest trends in business data
transformation?

The shift to a data-driven business

model – where decisions are made
based on what we know to be true rather
than on our intuition – is at the heart
of the wave of digital transformation
that has been sweeping through all
sectors in recent years. Data analysis
helps us respond with more certainty
in the face of uncertainty. When wars
and pandemics disrupt the order of
things and lead to massive inflation,
using advanced data analytics to
make better decisions becomes more
critical for businesses.

AI also needs data to learn, which

frequently means dealing with
sensitive personal data. For this reason,
it is imperative for organizations to
deploy responsible and trustworthy
AI. It is key to ensuring that the values
of inclusiveness and diversity are
respected. Data has an essential role
to play in creating a more ethical and
just world.

Another major emerging trend, linked

to the acceleration of global warming,
is the use of data as a tool to help
reduce ecological impact. Data and
algorithms can measure the carbon
footprint of activities and analyze
how to develop products, services
and infrastructures in more energy-
efficient ways, by identifying sources
of waste and inefficiency.

4
 INTERVIEW

Artefact identified two drivers: first,

to structurally reduce the carbon
emissions of logistics and digital
infrastructures, and second, to
engage consumers to participate
in this ecological transformation by
displaying the carbon content of their
shopping baskets and recommending
actions to compensate for their
emissions.

Business data maturity has

advanced rapidly over the past
decade. How has Artefact evolved
as a leader in data consulting
services?

Companies have implemented data

governance policies, which are a
prerequisite for any transformation,
but there are still sectors that lag
far behind in terms of their data

“
processing, with a real potential for
efficiency, such as healthcare or
“heavy” industries. This is especially
true in comparison to the consumer
and retail sectors, which have begun
their data revolution and which we
It is key to ensuring that the
know very well, such as L’Oréal,
Danone, Unilever, Samsung, etc.
values of inclusiveness and diversity
We started to transform marketing
are respected. Data has an
departments by making them more
profitable and relevant in their multi-
essential role to play in creating a
channel media investments with
pioneering targeting, measurement
more ethical and just world.”
and personalisation solutions. For
the past few years, we have also been
deploying acceleration programs in
all business areas (Sales, Supply
Chain, Operations, Call Centers, HR and services, and streamlining internal currently used by more than 10,000
& Finance, etc.). We create value operations to reduce costs and waste. Orange technicians throughout the
wherever there is data, and with our For example, we’ve been working with country – a resounding success!
clients, we improve their processes the Orange telecommunications
and produce customized business group for over six years, and among This case perfectly illustrates
applications. the many use cases for leveraging Artefact’s firm belief that to achieve
the company’s automation and AI true data maturity, companies have no
Can you give us a concrete potential, we deployed a solution choice but to make data accessible to
example that shows, from a precise with their teams to optimize their everyone: not only to experts, but also
business and operational objective, technicians’ interventions on the to operational staff in the field. This
how Artefact designs AI solutions fiber network. The solution is based will lead to new forms of augmented
that improve business uses? on visual recognition technology that work, where applications and their
helps operators improve the quality interfaces put intelligent information
Data is the key to understanding of their installations or repairs. This in everyone’s hands to work more
customers, developing better products application, available on a tablet, is efficiently and with more autonomy.

5
Listening to you, it’s clear that
data should no longer be a subject
reserved for experts only. How
does Artefact see the vision of data
democratization materializing?

The companies that will endure are

those that successfully foster a data
culture with access to knowledge and
data for all.

We’ve undertaken several initiatives

in this area that are highly strategic
to Artefact’s positioning as a major
player in data democratization, in order
to fully exploit its potential for positive
transformation.

“
The first was the launch of the Artefact
School of Data two years ago, a key
pillar in our strategy of providing
clients with training adapted to the
constantly evolving skills of the data
Democratizing data and
industry. We’re also developing “à la
carte” e-learning platforms for clients
making it accessible to all is key
to quickly share knowledge of data and
AI with all of their employees.
to accelerating business and
And at the beginning of this year, we
creating value.”
launched our Artefact Research Center
under the leadership of Emmanuel
Malherbe (class of 2008), to bridge the
gap between fundamental research After a year of robust growth in acquisitions in particular enabled
and its democratization for business 2022, what are your forecasts for us to expand Artefact’s portfolio of
applications, in collaboration with 2023? clients and services: the acquisition
clients who provide access to their of Startup Inside, a pioneer in open
data and use cases. To achieve this, we After +50% organic growth in 2022, innovation & intrapreneurship
have partnered with several renowned our objective is to maintain the strategy consulting and international
university laboratories, including momentum in 2023 with a sustained Data and AI conference organizer,
CMAP and MEI at Polytechnique, to recruitment policy in France and in and, more recently, the merger with
host PhD students at Artefact, who our 16 subsidiaries in Europe, MENA, the Arca Blanca group, a leader
work on data model improvement and Asia, North and South America. We’ve in data consulting in the United
organization studies. They will publish just deployed our Artefact Africa entity Kingdom.
scientific articles and participate in from Morocco and will soon open an
international conferences to share office in Korea. We will also accelerate We are quite optimistic about the
their knowledge. our development in LATAM and the future because even though the
United States. economy is currently strained,
These programs are just the first steps companies need to better understand
we’re taking to democratize data and Artefact’s expansion also involves the shifting environment and find
help our clients transform faster and an ambitious M&A policy that will rapid solutions for adaptation and
better. continue in 2023. In 2021, two progress through data.

6
 DATA FOR RETAIL

Becoming data-driven (and reaping

immediate benefits) has never been
easier in the retail sector

Data technologies are now at the head of the aisle in

the great supermarket of retail opportunities. Serve
yourself: take advantage of these unbeatable offers Jérôme Petit
today! Managing Partner
Retail & eCommerce
ARTEFACT

Par ticularly turbulent market Machine learning to optimize

conditions in 2022 disrupted the retail prices…
sector’s pricing and supply chain
policies. While price optimization Inflation has focused retailer
has always been a daily concern for attention on the daily challenge of
this industry, the 5.2% increase in price optimization: how to identify
consumer prices over the whole of products that are most sensitive to
2022 (compared to 1.6% in 2021) price variations (elasticity)? Which
and its acceleration in the first products shape the price image of
quarter of 2023 have intensified the a brand (known value items)? How
need to react. More frequent price to react in real time to competitor
adjustment decisions must be made repositioning? Since promotions
every day to absorb cost increases represent a growing share of sales
or react to competitor repositioning, (more than ever, consumers are
and be immediately deployed in store looking for bargains that fit a tight
networks. budget), how can they be optimized
without impacting profitability? All
To make matters worse, a shadow still the answers are in the data! Customer
hangs over supply chains, inherited data and proprietary sales history can
from pandemic events. Many retailers be combined with external data such
are no longer sure they can offer all as competitor price and promotion
products due to lack of stock. records, seasonality, calendar events,
and, yes, even the weather (!) to make
To address both of these challenges, the right decisions. Retailer pricers
players who are able to rely heavily on and category management teams
their data and process it instantly and usually do all this. Unfortunately, they
at scale through machine learning often need to work in Excel but have
stand out. Of course, you can’t limited time to devote to it, due to
go from data-rich to data-driven their other responsibilities.
overnight, but contrary to common
fears, the maturity of the technologies So, what’s new? Why, it’s the
is such that the first building blocks, availability today of proven off-the-
which can rapidly bear fruit, can be shelf algorithms that automate
deployed in mere weeks. the simplest decision-making

7
processes. It’s the seamless and need to be served as well. Once again, the transformation. Data is about
massive availability of third-party data science comes to the rescue of people. This is Artefact’s slogan, and
data in the market. It’s the ease retailers by allowing them to better rightly so.
of use of technology stacks that control inventory management. By
allow millions of transactions to be leveraging machine learning, retailers Directly monetizable data…
processed in a few milliseconds. can now analyze receipts in real But perhaps not immediately
Today, it only takes three months to time to immediately detect out-of- for everyone
build a data platform that combines stock items, calculate the spread of
all transaction data, promotions, uncertainties across all links in the The icing on the cake is that data itself
stock, product hierarchy, store chain to better size buffer stocks, is a goldmine, thanks to retail media
hierarchy, customer data, etc. And a or improve stock allocation under and data sharing. As digital signals
technology partner can manage the an infinite number of constraints become more difficult to capture, the
infrastructure, resource deployment (to optimize costs, shorten delivery billions of transactions and customer
and network dimensions in the cloud times, or reduce carbon footprint). interactions that retailers generate
through its managed services. Today, have become a critical strategic
like Monsieur Jourdain, if you know Democratizing data use advantage for them. This data, which
Excel and PowerPoint, you’re a data throughout the company: provides in-depth understanding of
analyst without realizing it: in a data is about people consumer expectations, has great
matter of days, you can take control potential for monetization. But it
of data in BigQuery (Google), Synapse In a business where margins are so represents a profound, existential
(Microsoft), or Snowflake and build tight that operational excellence is a transformation of the retailer business
interactive dashboards in Looker, necessity, the notion of a data-driven model: moving from a self-financed
Power BI or Tableau. company is far from new. What’s model (with negative working capital)
changing today is the ease of access but with very narrow margins, to a
…and better manage and use of technological platforms. model where initial investments are
inventory substantial but margins are high.
If technology is no longer a barrier, A Copernican revolution, perhaps
In recent years, the health and the challenge is still to make these not the easiest to undertake for all
geopolitical context has also solutions available to the widest players.
challenged supply chains. Today, possible audience. To democratize
the supplier service rate varies from their use, simple solutions need to In the great supermarket of retail
week to week and delivery times be deployed on a massive scale, value creation opportunities, data
can be very uncertain. Distribution employee training programs need technologies are now at the top of
channels have also grown highly to be multiplied, whether on-demand the shelf, in self-service. Retailers,
complex: not only do stores need or more intensive, and events (e.g., why wait to share these unbeatable
to be stocked, but home deliveries, hackathons) need to be organized to offers with your partners?
click-and-collect, and partnerships engage the managers who are driving

8
 DATA FOR RETAIL

CASE STUDY

CARREFOUR GROUP
How Data & AI can accelerate
sustainable business transformation

CHALLENGES

How Artefact is helping Carrefour achieve carbon neutrality for its

e-commerce activity?
Carrefour Group is the leading European retailer and Reducing the carbon emissions of its logistic and
the world’s second largest retailer, and is present in digital infrastructures;
more than 30 countries. Carrefour’s international Engaging customers to become participants in
profile raises a number of ecological challenges ecological transformation.
and a desire to offer its customers, regardless of
their level of awareness, quality food and services This aim also provides a triple opportunity for the
accessible to all. Group’s e-commerce activity: reducing its operating
costs, improving its NPS score (customer satisfaction
The Group’s aim is to become the world leader in food indicator) and anticipating legal changes.
system transformation for all by committing to four
major objectives, including achieving carbon neutrality To seize these opportunities and take concrete action
by 2030 for its e-commerce activities. on each of these levers, Carrefour must be able to
measure all greenhouse gas emissions by drawing
To achieve carbon neutrality, three main levers of on real data that compiles all data storage, transport
action have been identified: and logistics activities, from first click to final delivery,
Measuring the ecological impact of a delivery in order whether to the home or by store pickup.
to manage the strategy;

9
 SOLUTION

Defining a reliable, actionable, transparent carbon

measure for Carrefour and its customers

The first step for the Artefact and Carrefour teams The second step was to collect activity data in
was to agree on the scope of action for measuring this order to convert it into carbon emissions. As this
carbon footprint. They decided to limit themselves data wasn’t already present and documented in
to measuring greenhouse gas emissions generated Carrefour’s data platform, the business teams
by e-commerce orders in 2021 (logistics, warehouses, e-commerce) had to be
brought together to obtain it. This step proved to
be crucial to the operation’s success, as it allowed
all stakeholders to become ambassadors for the
“The challenge we gave Artefact was group’s “carbon neutrality 2030” objective.
to calculate the CO2 emission of an
Carrefour’s strategy for measuring its carbon
online order.How much CO2 will a footprint was based on a systemic, unifying, long-
customer produce if their order is term, iterative approach. The strategy was successful
delivered or if it’s picked up at the thanks to the participation of over 30 employees
store?” and the involvement of Carrefour customers via
their “Engaged Consumers Clubs”.
Bertrand Swiderski
“Today, we recognize that consumers are
Chief Sustainability Officer becoming experts on these topics. They want
CARREFOUR to understand how things are done and want to
challenge companies. Thanks to them, the project
has matured.”
Léonard Cahon
Consulting Manager - ARTEFACT
Encouraged by these initial results, Carrefour will
continue its commitment by publishing the carbon
footprint of each of its orders on its e-commerce
site in the near future.
“Soon, customers will clearly see the number of
kilograms per CO2 on their orders, thanks to the
insights gained from our carbon assessment.”
Manuel Chatain
E-commerce CSR Manager - CARREFOUR

10
 RESULTS

Opening a wider field of Using data as a key lever to

possibilities and ecological help businesses achieve their
alternatives environmental objectives
By analyzing its carbon footprint and implementing Carrefour’s e-commerce, supply chain and
this first measure, Carrefour now has a way to logistics platform teams worked together to meet
pilot its e-commerce carbon emissions reduction this challenge, supported by the collaboration
strategy. The group can now encourage its and expertise of the Artefact, Aktio and Google
customers to review their consumption patterns teams. The project is part of Artefact’s “Data for
in order to be more responsible, encourage its Sustainability” solutions, which aim to create a
service providers to reduce their emissions, and positive impact on the environment through data
also promote internal awareness by proposing by accelerating the ecological transformation of
several possible levers of action: businesses.
•Act on the choice of delivery slots in order to
optimize truck loading, routes and schedules; “We expected a very clear vision of what each basket
order would emit. Our request was complex, but
•Increase the number of clean vehicles (biogas, Artefact responded to it with great dynamism and
electric or hydrogen) by 2030; agility.”
•Reduce the amount of packaging used. Bertrand Swiderski
Chief Sustainability Officer - CARREFOUR
And to ensure the sustainability of this
measurement and its easy use by the teams,
Artefact teams worked on three elements: “At Artefact, we believe data will play
a major role in helping companies
•A dashboard to run trajectory simulations by
combining forecasted activity data; achieve their carbon neutrality goals.”
•Training to teach how to use and modify the Vincent Blaclard
dashboard; Partner - ARTEFACT
•Detailed documentation to enable employees
and clients to understand and reproduce the
measurement process from start to finish.

11
Data-driven marketing:
the rise of the Customer
Data Platform

Florian Thiebaut
Managing Partner - Data-Marketing
ARTEFACT

A game-changing technical
and legal environment
Following Safari’s lead in 2016,
Everything seems to justify the current explosion the world’s three main browsers
eliminated (or will eliminate) the use
of the Customer Data Platform (CDP) market. of third-party cookies. On the mobile/
CDPs’ main advantage over older generation Data tablet devices side, Apple’s iOS 14
Management Platforms (DMPs) is that they ea- now requires explicit consent for any
mobile ID collection.
sily integrate identifiable first-party data (email,
phone number) and aren’t dependent on using As for regulation, GDPR laws in Europe
third-party cookies or browsing data to refine have given consumers more control
over their personal data, requiring
customer and prospect knowledge. them to give explicit consent for
the use of cookies. This regulation
CDPs are a true asset in a world that is becoming represents a major shift in the world
increasingly cookie- and ad ID-free. At a time of data-driven marketing, as it has
when the pandemic is forcing brands to digitise reduced the number of cookies
placed on European devices by 30%.
at breakneck speed, and when the transforma-
tion of the technical and regulatory environment This global trend restricting the use of
surrounding advertising trackers is forcing data IDs and advertising cookies sharply
impacts the targeting capabilities of
marketers to revise their approaches, CDPs are advertisers, who are often dependent
here to optimise the customer experience. on third party data. The vast majority

12
 DATA FOR RETAIL

of them use or have used retargeting a CDP environment based on a suite This data completes a database that
and old generation DMPs that rely of tools that is both compliant and is incomplete at certain points in the
heavily on segments fed by third party sustainable. This will enable data consumer journey. Examples might
data. to be collected, stored, processed, be an agreement between an FMCG
visualised and activated, whatever brand and a retailer, a mobile phone
Along with targeting, measurement the source. From this foundation, the manufacturer with a telco or a hotel
must also be transformed. With focus must be on first-party data. chain with an airline.
more stringent consent collection
requirements, it’s more difficult to D ata governance: Brands need Three types of data to
collect the consumer IDs needed to to rethink data governance and activate via a suite of tools
track impressions, clicks or views, processes to enable secure and
and reconstruct complete customer compliant end-to-end data collection. First- and second-party data are key
journeys. to meeting the challenges of the post-
Audience segmentation: This data, cookie world. But what are they and
Four pillars for a sustainable centralised for a unified view of the what tools can be used to manage
data strategy consumer, can then be used to create them?
new audience segments and define
To maintain the same performance new metrics for measuring campaign PII or Personally Identifiable
and differentiate themselves from the results. Information is essentially CRM
competition, advertisers must design (customer relationship management)
a sustainable data strategy and S econd-party partnerships: In data. It can precisely identify an
exploit their customer and prospect addition, it’s becoming increasingly individual and is often an email address
data to its full potential. strategic to form so-called “second or a phone number for example.
party” partnerships with other partner Once anonymised, it can be used
This requires focus on four actions: companies to exploit first-party data via the APIs of media partners (e.g.,
and create win-win situations. Google Customer Match, Facebook
The CDP: The first step is to establish Custom Audience/conversion API,

13
 Amazon, WeChat, etc.) to build relying on third party data
audience segments, perform audience
extensions, and reconstruct paths Data that is purely media-related,
to measure the influence of digital such as campaign impressions,
campaigns on offline sales, etc. video views and click rates, is more
voluminous and less granular than
Non-PII data can be browsing data the other two types of data. It is
that cannot lead directly to the more difficult to use but there is a
identification of an individual. It robust market of tools capable of
can be used to build more granular treating it in a secure and compliant
segments via analytics and audience manner, such as Google Ads Data
creation solutions for measuring Hub, Facebook Advanced Analytics
precision marketing actions without and Amazon Marketing Cloud.

These different data flows are

injected into an ecosystem of
interconnected tools, which are
useful for a range of tasks — from
data collection to performance
measurement of the actions carried
out — and can be activated on all
channels, whether media, direct
marketing or site personalisation.
This entire ecosystem, the result of
all the connections built between the
different tools already used by the
company (also known as “full-stack”
solutions), is what is called the CDP.

When it comes to the adoption of this

way of working, the numbers don’t
lie. Fundraising for CDP providers
is soaring, the tech giants are all
positioned in the sector, and the
number of users is exploding.

In fact, according to the Customer

Data Institute, the market increased
30% from $1 billion in 2019 to $1.3
billion last year. Estimates see this
figure reaching $1.55 billion in 2021 as
conditions are even more favourable
for the adoption of CDPs.

As the data-driven world continues

to evolve at a rapid pace, there
seems little doubt in the business
value of the CDP. Now is the time for
organisations to consider deploying
this future-facing technology.

14
 CASE STUDY

CONFORAMA
AI-enabled personalization boosts
Conforama CRM campaign revenues
CHALLENGES

Conforama is the second largest home furnishings Several challenges needed to be addressed through
retailer in France and is present in seven countries, this use case:
with 300 stores, including 200 in France. The
company sells furniture and decorative items in kit • How to understand the needs of three million
form and posted sales of 1.7 billion euros in 2022. customers and recommend the most relevant
products from a catalog with 42,000 references?
As a gateway brand, Conforama’s goal is to “Make
what people want most accessible at the best price.” • How to propose only products currently in stock,
It’s an ambition backed by a transformation plan to on promotion, and not already suggested to
deliver an omnichannel experience through data and customers?
AI. An initial audit and data marketing vision with
Artefact identified and prioritized 12 use cases and • How to easily operate and maintain the technical
25 technical and organizational enablers. The first solution?
use case was to integrate a personalized product
recommendation into the company’s weekly emails.

15
 “Time savings, yes, but above
all a business benefit for our
CRM teams. Because thanks to
this personalization, customers
click more and therefore buy
more. We’ve gained 15% of
the click rate following the
personalization of these emails,
which represents several million
in incremental sales.”

Mélodie Charles, Marketing Director

CONFORAMA

SOLUTION

Saving consumers time, improving business productivity

By using machine learning algorithms to analyze In addition, AI-based product recommendation
user data, such as preferences, purchase history and can boost business productivity: machine learning
online behavior, artificial intelligence-based product algorithms can analyze large amounts of data in real
recommendation suggests products relevant to time, allowing companies to continuously monitor
consumers in a personalized way. This allows customer trends and buying behavior. This can
companies to better understand their customers’ help organizations better understand customer
needs and recommend products that match their desires and quickly adapt their product offerings
interests, resulting in increased sales and customer accordingly. It can also enable companies to optimize
retention. their inventory by offering products that are more
likely to sell, which can lower costs and maximize
One of the main benefits of this solution is that it profits.
saves customers time. Rather than scrolling through
countless product pages to find what they’re looking Lastly, AI-based product recommendation can offer
for, customers can quickly access a selection of significant business benefits. By suggesting relevant
recommended products that specifically meet and personalized products to customers, companies
their needs. AI-based product recommendation can improve their conversion rate, increase sales
can enhance the online shopping experience and and strengthen their brand image. From a market
encourage customers to return for more purchases. perspective, AI-based product recommendation has
A strategic advantage, given that 72% of consumers been shown to deliver +2.5% incremental growth.
only interact with marketing messages that are
personalized and tailored to their interests.

16
 DATA FOR RETAIL

A first use case focused on email campaign

personalization
Prior to this project, all Conforama customers • Building the “Collaborative Filtering” model to
received emails featuring the same eight products calculate customer appetite for the product catalog;
selected each week by the marketing teams. This
was a labor-intensive task, as it required identifying • Product filtering based on available inventory,
the eight products most likely to interest three million commercial news (sales, promotions, etc.), past
customers, each of whom had unique interests. activations and purchases;
All this time spent analyzing data could have been
spent on more strategic activities, such as creating •
P roduct data enrichment (photos, prices,
editorial content for those emails. descriptions, etc.) for activation.

Today, an email is sent to every Conforama This solution is based on 16 data tables, 25
customer each Tuesday containing eight product transformation and modeling steps, and 40
recommendations. But these recommendations automated quality tests. Dozens of iterations of
are personalized according to purchase history, and the model made it possible to choose the most
filtered exclusively for products that are on sale, are efficient approach based on transaction history.
available in stores, and that haven’t been featured Thanks to this solution, Conforama now generates
in previous activations. several million recommendations each week in 45
minutes at a cost of 50 euros per week.
The implemented AI solution includes 4 main data
processing steps: In other words, if you count development and
operation costs, as well as incremental sales, the
• Collection of transaction histories, customer and project break-even point is reached in one week,
product references, then data preparation; with an automated and reliable solution.

17
 RESULTS

A smooth transition to AI:

lessons from Conforama’s success story

For many players, there are three challenges linked • Select a first use case and functionalities that can
to their level of maturity: be quickly implemented and measured to put the
organization on the road to success. For example,
LEVEL 1 this initial victory means Conforama can now plan
the deployment of product recommendations in
Personalizing a currently rule-based touchpoint stores or the improvement of their algorithm thanks
using an AI algorithmic approach; to browsing data.
LEVEL 2 • Ensure the data is reliable. Good data modeling
relies first and foremost on good quality data. For
Extending AI-based personalized recommendation Conforama, exploratory analyses were performed
across the entire customer journey (similar products on more than 50 tables to select data sources
/ complementary products / suggestion based on in areas such as customer knowledge, product
purchase history); repositories and transactions.
LEVEL 3 • Use technologies that allow teams to deploy a
technical solution quickly and collaboratively.
Optimizing the orchestration of recommendations
Conforama selected the most appropriate tools
across channels to ensure an omnichannel
for this type of workflow: DBT, BigQuery ML and
experience.
Vertex AI for their performance, modularity and
Level 1 is often the most difficult, as it requires laying portability.
the foundations for four separate dimensions: target
• Build a dedicated team capable of dealing with
vision, user experience and priorities; data sources;
all potential problems, and adopt a test and learn
technological tools; project team and work method.
approach. To do this, a multidisciplinary IT /
The Conforama example offers valuable lessons Conforama business team was formed, and a
about these four dimensions: 2-week sprint approach was adopted.

18
 DATA FOR RETAIL

Retail Media:
An indispensable
asset for brands
While Retail Media represented only 9% of
digital media investments for brands in 2019, it
will soar to 43% of these investments in 2023
and is expected to double in value by 2024
to reach €100bn. Vincent Cailliot, Director
of Data Consulting and Sidney Zeder, Senior
Consulting Manager – Data Marketing, both
of Artefact, explore the opportunities of retail
media for Consumer Packaged Goods (CPG)
brands.

Retail Media investment As a result, brands need to find new

driven by regulatory and tech data sources to build their digital
activation strategy. One solution is
developments
for brands to better leverage their
The increased importance of Retail first-party (1P) data assets, i.e.,
Media in the digital strategies their proprietary data, by collecting
of brands can be explained by more and better data from their
the evolution of the availability of customers. But to create digital
Vincent Cailliot consumer data, which is at the heart activation strategies with long-term
Director Data Consulting of any media personalization strategy. value, proprietary data is no longer
ARTEFACT sufficient: it needs to be enriched
In the past, third-party cookies were with data – especially transactional
mainly used to drive this strategy. data – from retail partners.
Regulatory developments to better
protect consumer data, including Facilitated data sharing between
the GDPR in Europe, have led to new brands and retailers allows brands
technological constraints, making to deploy ever more relevant digital
cookies obsolete: This process began marketing strategies with a high
as early as 2016, when Safari removed level of autonomy. This increased
cookies from its platform. It continues collaboration and data sharing
today with Google Chrome, the most between retailers and brands is
Sidney Zeder popular browser, announcing the possible thanks to “data clean rooms”
Senior Consulting Manager removal of third-party cookies from such as Amazon Marketing Cloud,
ARTEFACT its platform next year. LiveRamp or Decentriq, which allow

19
the sharing of personally identifiable to transactional data to build their sales and ROI, enabling effective
consumer data at the individual level Retail Media strategy. While the optimization of activation plans.
in an anonymous way. majority of retailers in the US have
launched Retail Media offerings, Retail Media is just the next step
A rapidly evolving ecosystem most retailers in Europe are still in the towards more collaboration
experimentation and use case-testing between retailers and brands. In a
The ecosystem of technology phases; few have yet industrialized long-term partnership perspective,
partners around Retail Media is highly use cases with brands. collaboration and data sharing can
fragmented and constantly evolving, enable the implementation of more
with partners that are more or less Valuable use cases beyond advanced category management
specialized depending on major Retail Media and supply chain use cases, such
Retail Media activities: first-party, as the analysis of the long-term
second-party or third-party cookie LeRetail Media allows brands to value of existing promotions or
data collection tools, data processing address marketing use cases from the prediction of in-store product
and audience creation, activation consumer insight generation to digital demand and thusoptimize supply
or analysis, etc. The challenge for campaign activation and marketing
brands will be to identify which performance measurement. The chain operations.
combination of technology partners availability of transactional data
will best meet their needs, depending (previously unavailable to CPG Which Go-To-Market strategy
on their current technical ecosystem B2B2C brands) at the “individual” level to launch?
and their own business challenges. enables the construction of insights
and activation plans that are all the For retailers, it’s important to define
Retailers are also an essential part more impactful. The same data can a new offer to monetize their data.
of this ecosystem, providing access be used to measure their effect on This can range from monetizing their
owned media inventory (website), to
sharing data “as a service” in a clean
room, to offering services (campaign
management or reporting as proposed
by Amazon for example). These new
offers can be marketed internally or
via partners. The internal or external
development strategy will determine
the associated costs, in terms of
salaried resources (commercial and
technical profiles to be recruited)
and technical resources (clean room
tools, technical infrastructure to be
set up).

For brands, the approach will be

more traditional, from defining their
business needs by identifying and
prioritizing use cases, to setting
up the partnership with their key
business partners. Likewise, the
implementation of pilots with a
retailer to test the value of use cases
can be carried out by a third-party
partner.

20
 DATA FOR RETAIL

CASE STUDY

UNILEVER
How does Artefact support Unilever
on Retail Media use cases to increase
its sales?
CHALLENGES

Thanks to Retail Media, Unilever identifies new growth

opportunities and increases sales of its priority product
categories.
The global Unilever group has a portfolio of 400 offers were born, i.e. 2nd party data partnerships.
brands that are anchored in the daily lives of its 5 A retailer’s data is made available to a brand to
billion consumers. Present in France for 125 years, enrich its own data assets in a win/win approach.
Unilever is the leader in several market categories This movement was initiated in 2012/14 by pure
including ice cream, condiments and deodorants. players such as Amazon and Alibaba, and traditional
The consumer packaged goods (CPG) sector has players such as Casino, Walmart and Carrefour have
the particularity of being intermediated. Since gradually followed.
the distribution of their products is carried out by Accompanied by Artefact since 2019 on various data
different retailers, there is little direct relationship issues, the Unilever Group seized this opportunity
between CPG brands and their clients. Yet, consumer to identify new growth levers, develop a common
knowledge is key to optimize media and promotional consumer and product approach with a leading
strategies, product assortment in the territory or to retailer and increase sales of certain priority product
identify new growth levers. categories. This is the case, in particular, for the
It is from this challenge that the so-called Retail Media Magnum brand’s mini ice cream bars range.

21
 “Retail Media is a win/win strategy for brands and
retailers. Retailers’ data allows us to enrich the
shopper’s knowledge and accurately measure
our activities on all channels, throughout the
transformation tunnel. For their part, retailers find
a new source of revenue and differentiation from
their competitors. In addition, it is a way to better
satisfy their clients with more personalized offers
and a better anticipation of stock shortages.”
Sarah Baqa
Head of Performance Marketing - Unilever

Retail Media in a win/win Retail Media infuses

partnership with Trade Unilever’s media,
Marketing promotional and supply
Advertising expenditure (media and traditional chain strategy
marketing) in the consumer goods sector amounts to
nearly 680 billion dollars worldwide and 16 billion in Unilever, together with Artefact, has identified 6
France. The investments allocated to Trade Marketing strategic axes to exploit the full potential of Retail
are just as massive. Media:
While the advertising levers have been highly digitized • Media activation
over the last twenty years, Trade Marketing is still very • Measuring Customer Lifetime Value
underdeveloped. However, driven by e-commerce,
• Coupon customization
it represents half of the budgets of CPG advertisers
(600 billion dollars worldwide, 16 billion in France). • Optimization of store assortments
Trade marketing is strategic for these brands because • Short-term sales forecast
it consists in carrying out actions in partnership
with mass retailers to better meet consumers’ • Supply of shelves
expectations: prospectus, merchandising, point of Since the CPG group initiated these retail partnerships,
sale animations but also sales forecast and supply each of these axes has been worked on through
chain optimization. Thus, the digitalization of trade concrete use cases. For this, Unilever has benefited
marketing represents a real growth opportunity for from Artefact’s end-to-end support. This concerns
advertisers and brands. data strategy, the launch of pilots, the construction
Retail Media is based on this concept of partnership of new data products, but also the provision of
in the promotion of brands’ products. The sharing of human resources (data analysts, data scientists and
2nd party data favors client knowledge and allows a data engineers), as well as the training of Unilever
better segmentation and therefore a better targeting employees.
of advertisements or promotions. This targeting can
be activated in both proprietary brand environments
and in external audience extension environments.

22
 DATA FOR RETAIL

SOLUTION

An iterative approach ”Our data strategy is really

ambitious, so we wanted
The Retail Media project was envisaged through an
iterative approach in 4 steps: to be accompanied by a
• Identify and select the right distribution partners team of experts. We chose
based on their pre-existing relationship with Unilever Artefact, which already
and their technology infrastructure. has experience with similar
• Laying the foundations of the partnership, i.e.
the use cases that can be implemented as well as
products and works in a
the necessary legal and technical requirements. very advanced way on data
The modus operandi is also built to facilitate and Retail Media solutions.
collaboration between all stakeholders (agencies,
internal and external teams of Unilever and the
In addition, they have the
brand). technical capabilities and
• Implement pilots for each use case. They allow for resources to take use cases
testing and optimization of devices on key Unilever to scale. This is the key to
brands (Magnum and Sun) before scaling up.
this long-term partnership.”
• Deploy devices on a greater number of brands,
use cases, markets and partner brands.
The vision of this data partnership is therefore based Sarah Baqa
on a virtuous circle: the measurement of actions
carried out makes it possible to collect insights
which in turn feed the next activations.

23
 DATA FOR E-COMMERCE & RETAIL

”Retail Media allows us to activate audiences

and precisely measure the link between digital
actions (advertising, promotions) and sales. The
data from the retailer’s loyalty cards also has
the advantage of tracking consumers over time
(Customer Lifetime Value) to see if they have
increased their purchase frequency, if they have
increased the value of their basket or if they have
switched brands, etc.”
Florian Thiebaut,
Managing Partner – Data Marketing - Artefact

RESULTS

A use case conducted on

Magnum’s mini bars segment
For example, retailers’ data was activated in the media
to meet a challenge for the Magnum brand: to confirm
its leadership during the summer period, which is key
for this product line, and to acquire new clients in the
Confectionery segment (mini bars).
To achieve this, the first step was to use shopper
data to identify the different consumer segments.
We showed them the same banners during a defined
period. This allowed us to measure the increase in on
and offline sales, but also to track the buyers recruited
to retain them post-campaign. Through this first wave,
Unilever was able to identify the best performing and
most profitable audiences. The ultimate goal is to use
these results at scale to target these audiences with
more personalized ad creative and promotional offers.
This pilot campaign significantly increased the brand’s
revenue.
Unilever’s data and Retail Media strategy will continue
in 2022 alongside Artefact. In addition to the two
«Precision Media» levers, the exploitation of the 4
other identified work axes is also planned.

24
 DATA FOR RETAIL

Data monetization opportunities for

retailers: Retail Media within the
CPG/Retailer data ecosystem
Retail media has been around for quite a while, but
thanks to the evolution of new uses of consumer data,
its potential is gaining attention. Sidney Zeder, Senior
Manager and Gaétan Bélan, Senior Data Consultant
& Product Owner, both of Artefact, explore the
opportunities of data monetization for retailers.

Retail media has been on the rise said they were already investing in
on digital platforms for the last six at least one retail media platform.
years, most notably on Amazon. The This represents approximately 17%
Covid-19 crisis accelerated this trend of digital budgets already allocated
for traditional retailers. Retail media, to retail media.
in simple terms, is the means for
retailers to sell media inventory on Media investment has effectively
their e-commerce platforms. Because shifted down the marketing funnel.
Gaétan Bélan the Covid-19 pandemic fueled the Although many brands’ search
Senior Data Consultant shift to digital ways of buying, such investments still flow into the “Google
& Product Owner as e-commerce or click-and-collect, family,” we’re seeing brands diversify
ARTEFACT even for grocery shopping, retailers their digital spend into e-commerce
had no choice but to go with the flow. platforms to capitalize on the “search
destination” status they hold. When
In fact, between 2019 and 2020, CPG you’re on Amazon as a consumer,
e-commerce penetration increased you’re very close to the “moment of
by five points, from 10 to 15 percent. truth”: you’re in a purchasing mindset.
For retailers, the downside is that Therefore, when you’re on Amazon as
margins are lower in e-commerce a brand or product, the closer you can
than in brick-and-mortar. The upside get to that funnel, the better. Goldman
is that by selling online, they collect Sachs expects this trend to translate
a lot of consumer data that can be into a 6-8 percent increase in total
Sidney Zeder monetized or used to create new CPG e-commerce sales through retail
Senior Consulting Manager services. In a Goldman Sachs study, media over the next four years.
ARTEFACT 82% of CPG companies surveyed

25
Retailer data monetization behavioral (e.g. what did they look retailer to build advanced audiences
opportunities with CPG at), loyalty data (e.g. did they buy for a digital marketing campaign.
brands again), etc. This data is shared “as-a- Using the retailer’s transactional
service” in a data clean room where data, the brand was able to build
This close-to-the-funnel media brands can access the retailer’s data and activate two audiences: the
investment trend has created in a secure environment to carry out brand’s current ice cream buyers and
opportunities for retailers around specific use cases defined by the ice cream buyers from competitor
three types of data monetization with two partners. brands. As a result, the brand was
CPG brands: able to increase the uplift of its
Carrefour, for example, has created campaigns by targeting the two
1. Inventory monetization: traditional a consumer intelligence service relevant audiences with adapted
retail media consisting of selling called Carrefour Links, based on the messaging.
media inventory on proprietary LiveRamp clean room, where partners
assets. This can be offline inventory can access their cardholder data. This •Trade: the data shared by retailers
– retailers have long monetized is a self-service platform that allows allows brands to perform revenue
their customers to offer coupons or users to perform basic activities such growth management use cases by
specific promotions to brands in their as reconciling retailer and brand better optimizing promotions or
stores – but also their online inventory databases on individual customers assortment… It also unlocks store
on their own platforms, such as their to build a more complete view of the optimization use cases through
e-commerce website, where brands consumer and thus improve their enhanced in-store experience
can display banner ads, emails or experience. It also provides analytics or sales force optimization. For
even shopping mobile applications and measurement capabilities that example, one brand worked with a
to deliver personalized promotions Carrefour can bill to its partners. retailer to analyze the short- and
to their customers. long-term impact of promotions on
Access to this data can unlock three incremental margin. This enabled
2. Data monetization: retailers are types of use cases for brands: them to identify certain types of
monetizing existing consumer data promotions that were margin-
across CPG brands to support their •
M arketing: the data shared by destroying for both the brand and
customer centricity. 1P data shared retailers allows brands to gain the retailer, as opposed to those
by retailers originates from their insights about their consumers, that generated a positive long-term
loyalty program. The cardholder data activate them with media, or measure business impact.
they share can be socio-demographic marketing performance through
(e.g. the age of their consumers), transactional data. For example, an •O perations: the data shared by
transactional (e.g. what did they buy), ice cream brand partnered with a retailers allows brands to optimize

26
 DATA FOR RETAIL

their supply chain through inventory to brands, while more

demand forecasting and demand mature retailers are focusing on
management use cases. It can data monetization or even service
also fuel sustainability as well as monetization for the best-in-class
production and innovation use retailers.
cases.
Data monetization offerings require
Brands usually start implementing setting up technical capabilities,
marketing use cases to deliver short- such as a technical infrastructure
term value with a simple set-up, while to collect, store and process the first-
long-term partnerships can then party data to be shared, as well as a
address very valuable use cases for clean room to share data with brands,
Trade and operations that benefit both or even partnerships with DSPs to
partners. Across all three categories, enable brands to directly activate
data monetization unlocks better audiences created in clean rooms.
measurement capabilities: optimized But the business opportunity is worth
media performances, customer it: selling data as-a-service delivers
lifetime value calculation, dynamic margins often in excess of 80%,
budget allocation and global ROI compared to inventory monetization
optimization. where margins typically only reach
40%, as inventory assets are limited
3. Service monetization: along with and therefore less scalable.
inventory and data monetization,
retailers can unlock additional
revenue by providing different levels
of services to brands.

In the most developed additional

service offering, retailers can propose
managed services to brands, based
on shared SLA and KPIs. In a long-
term partnership approach, retailers
optimize their revenue potential
through best-in-class services
for brands. Amazon offers its key
clients advanced services such as
category management studies and
dashboards or MMM (Media Mix
Modeling) studies to help them
improve their strategy in relation to
this retailer.Des cas d’usages à valeur,
au delà du retail media
Data monetization: maturity
in the Retail Media market

Artefact has benchmarked over

20 retailers in the US Retail Media
market and analyzed their maturity
with respect to Retail Media based
on their value proposition and
the comprehensiveness of their
capabilities. We found that most
retailers in the US have launched
a Retail Media offering but are still
in a nascent stage, mainly offering

27
CASE STUDY CHALLENGES

CARREFOUR Retail ecosystem is

LINKS currently going through
a revolution
With Artefact are The old rules of retail have been deeply impacted

helping brands by two new factors: digital assets (digital media)

and shopper data, that helps leverage knowledge

to improve media of customer behavior and granular activation of

digital assets.

targeting thanks Retail media mixes digital assets and shopper

data. For many years, retail media has been

to Carrefour’s data helping brands and manufacturers become more

efficient in four different ways:
• customer insights: a deeper understanding of
your customers thanks to data from identified
members of customer loyalty programs
• precise targeting: moving from mass marketing
towards 1:1 marketing
• activation: omnichannel activations that can be
synced through several touchpoints enhanced
by shopper data
• measuring: measure of the return on ad spend
and impact on sales

28
 DATA FOR RETAIL

SOLUTION

“Retail media is an ecosystem

Carrefour’s vision on that needs a certain number of
their Retail Media mandatories on the retailer’s and
system manufacturer’s sides“
Thomas Faure,
Carrefour provides manufacturers with cutting-
Senior Consulting Manager
edge solutions to conduct retail media campaigns
and to develop their sales and display presence E-Retail Lead - Artefact
both in store and on their ecommerce website.
According to Nicolas Trolé, Carrefour Media has
the unique opportunity to be able to leverage
assets from the whole entire division, such as
5000 physical stores, a digital platform that
reaches 10 millions customers every month, and RESULTS
the key element in their omnichannel offering,
their unique data. Artefact as a
In 2021, Carrefour took the decision to develop a
new full stack platform on carrefour.fr, with Criteo trusted partner to
as partner, to offer advertisers new activation
solutions and better measuring and targeting.
maximise the value
of Carrefour’s data
Thomas Faure reviews the four pillars of
the partnership between Carrefour Links
and Artefact:
• Business consultancy: Manufacturers
“When we talk about data at want to understand where growth comes
Carrefour, we are talking about from. “Carrefour’s data helps brands
data from our customer loyalty identify business opportunities for them
to accelerate”
program, which represents 14
million cardholders, 7 million • Profiling and customer analysis: “Who is
my customer ?” One of the main challenges
customers registered online and for manufacturers is to define who their
offline, and 5 million digital-only customer is. Carrefour’s data helps brands
customers“ represent the personas of their brand
buyers
• Strategy, operation and activation: helping
Nicolas Trolé, brands create media activation strategies
Chief Revenue Officer - Carrefour Links (budget, KPIs)
• Advanced data use cases: projects that
take into account the whole retail value
proposition

29
 CASE STUDY

How Artefact helped boost Mattel’s online

sales on Cdiscount (RelevanC Advertising)
retail media platform?

CHALLENGES

Increase presence and visibility for the brands of the

Mattel Group on e-retailers media ecosystems by
leveraging retail media.
E-retail is a strategic distribution channel for the • in-platform, products that maximise the presence
Mattel Group as 32% of transactions in the gaming of products directly on the eCommerce website
category were conducted online during the 1st such as sponsored brands, sponsored products,
semester of 2020. merchandising banners
Retail media is used in an objective of performance • off-platform, solutions that generate additional
enabled by leveraging shopper data. Retail media traffic redirected to the product pages on the
is especially helpful for the conversion part of the eCommerce website such as search engine ads,
sales funnel. social ads, display & video ads
Retail media lives in different types of environments
such as:

30
 DATA FOR RETAIL

SOLUTION

High touch handling of Mattel’s retail media investments

in partnership with relevanC Advertising.

In 2020, Cdiscount cemented its position as the In addition to leveraging the e-retailer ecosystem,
number one French e-retailer, generating 25 million Artefact kept looking for ways to innovate and test new
unique visitors and 10 million customers (adding a solutions to build more expertise and scale projects.
million new buyers during the year). In that way, Artefact conducted off-platform campaigns
Cdiscount’s vision of its retail media ecosystem through Shopping for Partners solutions (Google for
covers the entire sales funnel and encompasses Retail) that were activated to support milestones,
different steps (brand awareness, consideration, such as product launches.
traffic, acquisition, and insights)
In February 2020, relevanC introduced a self-serve
platform, relevanC Advertising Platform, built entirely
in-house, that offers retail media solutions through
search and display on Cdiscount.
“relevanC Advertising is a best-in-class tool to operate
and manage retail media campaigns,”
says Maïana Darmendrail – Digital Manager & E-Retail
Manager, Mattel France RESULTS
“CDiscount possesses very mature retail media
solutions in search, display and video,”
states Thomas Faure – E-Retail Lead, Artefact
Ad campaigns six times
Artefact’s made the decision to perform an all-year more effective!
round online campaign aimed towards buyer intent.
Artefact defined custom audiences out of precise Mattel achieved extremely positive results
shopper data items (purchase history, buyer intent, through retail media activations in 2020:
search history, browsing history, income level ..). • 600% increase on return on ad spend for
The key to increase performance of media investments media investments using retail media
was to conduct continuous optimizations on both data (compared to traditional campaigns)
retail and media KPIs. • from 4 up to 10 euros earned when
“We have the intimate conviction that continuous spending one euro on display and search
optimization brings performance improvements,” ad campaigns
explains Cédric Chamoux – Directeur Retail Media, “relevanC Advertising has the most
relevanC Advertising qualitative data that you can find on the
• Optimizations were made on several factors: market. The daily optimization by Artefact
channels (onsite, offsite, search, display), segments really moved the needle,”
(audience, keywords), creatives, formats.
mentions Maiana Darmendrail – Digital
• Optimizations were based on several metrics: media Manager & E-Retail Manager, Mattel France
metrics, business metrics, retail metrics (stock level,
promotions, organic positioning) Brands are able to find success when mixing
business metrics with media expertise, i.e.
• Optimizations happened on different solutions:
sync retail signals (stock level, promotion
search advertising, daily improvements on
level) with media KPIs to better optimize
keyword selection, display advertising, bi-weekly
improvements on impact measurement advertising campaigns.

31
 CASE STUDY

CARREFOUR GROUP
How to reduce food waste
in the bakery-pastry department?
CHALLENGES

How to reduce food waste in the Bakery-Pastry

department?
Predicting demand as accurately as possible is one of lot of frustration among customers when they are out
the foundations of the retail business. However, this of stock at the end of the day. This is why Carrefour
challenge is becoming more complex as consumer Group, together with Artefact, sought to use machine
habits evolve. From now on, it is necessary to take learning and data science to optimize the prediction
into account the combined use of different sales of sales in the bakery-pastry department of its
channels (from e-commerce, to local stores, to supermarkets.
supermarkets) and the increasing demand for social This project meets a double objective: produce
and environmental responsibility. At the top of the list enough to meet the demand, while reducing
of CSR issues linked to mass retailing: food waste. “scrappage”, i.e. the volume of unsold products.
In France, bread is the third most wasted product This use case’s success relies, in particular, on
after fruits and vegetables. Indeed, bread, pastries the teams involving both the “jobs” concerned
and cakes are fresh products that have a very limited (department managers) and the various technical
shelf life. But, they are also items that can generate a profiles of Artefact and Carrefour Group.

32
 DATA FOR RETAIL

SOLUTION

A new algorithm to prevent For example, the algorithm’s recommendations were

first tested in pilot stores, only on pastries, to get
“scrappage” in the bakery- feedback from the field teams. Their comments were
pastry department used to improve the models, before the solution was
deployed across all bakery and pastry products in
the supermarkets.
In the retail world, every day is different. Sales are
highly dependent on the context: holidays, weather,
current promotions, merchandising highlights, etc. Easy-to-access
To take into account all the variables that have
implications on demand, we need to be able to and interoperable tools
analyze the petabytes of data generated by the
All of Carrefour’s data-driven use cases are fed by
billions of transactions carried out every year at
a centralized platform in the cloud, which makes
Carrefour, to which we need to add external data
the data accessible, formatted and documented.
that influences consumption. These calculations
The results of the processing are then fed back into
are only possible with Artificial Intelligence.
Carrefour’s information systems, shared by a wide
The Carrefour Group and Artefact teams thus started range of employees. Several teams, made up of data
with data from the sales receipts generated by more scientists, data engineers, but also data translators
than 200 supermarkets in France. Every day, this (pivotal profiles, acting as a link between business
data is collected, cleaned and enriched with external and data), are likely to consult, transform and process
sources – such as calendar data, for example – them for specific uses.
to build a sales history over several years. This
represents thousands of configurations for a single
day, depending on the assortment, product prices,
promotions, etc.
This data is used to train supervised machine learning
models, built on the basis of decision trees, which
determine the relationships between the target
variable (future sales) for each product and the
explanatory variables (promotion, cannibalization,
etc.).

Close collaboration
with the “field” teams
From an organizational point of view, the project was
led by multidisciplinary teams. Two teams on the
Artefact and Carrefour sides combined technical and
business profiles. The operational skills of Carrefour’s
retail professionals played a crucial role. They were
able to explain their business, their needs, and bring
their vision, in order to guarantee the success and
adoption of the solution “in real life”.

33
 RESULTS

Concrete results on the

number of discarded
products…and on the
brand image
A few months after the implementation of
this new prediction model, the results are
very positive.
In fact, over the last five months of 2021,
approximately 100 tons of pastries were
saved. At the same time, sales have increased
due to fewer shortages at the end of the day.
Finally, the Net Promoter Score, a performance
indicator closely monitored by Carrefour, has
evolved very positively.

Carrefour Group is
multiplying AI use cases
to improve customer
experience
For the group, future experiments follow the
same model: responding to business needs,
working jointly with operational teams, to feed
the customer experience. This acceleration
of Carrefour’s digital transformation was
made possible by the creation of complete
and expert data teams within the company, In fact, over the
and the deployment of data platforms in all last five months of 2021,
countries where the group operates. approximately 100 tons
The volume and wealth of data collected
by Carrefour provides a unique opportunity
of pastries were saved. At
to explore the major challenges facing the the same time, sales have
retail sector: omnichannel, e-commerce, increased due to fewer
anticipation of consumer habits, etc.
Carrefour recently unveiled other examples
shortages at the end
of how data can be used to improve the of the day.
customer experience: five-minute shopping
on Carrefour.fr, the implementation of
personalized assortments for local stores,
and the personalization of promotions.

34
 DATA FOR RETAIL

Demand forecasting: Retailers main challenge

Using Machine
is to go beyond past sales
to predict future sales

Learning to predict
accurately
Managing orders and inventory is

sales in retail.
the one of the strongest competitive
advantages that can help retailers
achieve success. And it is a real
challenge to master as it involves
processing a huge number of SKUs -
some of them that are even perishable
- ordered daily. We estimate that bad
Jérôme Petit inventory management, whether it’s
Managing Partner out of order items or excess stock,
Retail & eCommerce cost US retailers close to two billion
ARTEFACT dollars per year. For decades, retailers
have been relying on the analysis of
their past sales for their Enterprise
Resource Planning (ERP) that helps
them reduce their investment and
All industries aim to manufacture just the right exploitation costs. However, these
methods are heavily biased and are
amount of products at the right time. But, for not that useful when trying to predict
retailers, this issue is even more important as they accurate sales.
also need to manage their stocks efficiently. Too Dozens of signals to take
many items in stock is bad. Too few items in stock into account when assessing
is also bad. And to predict sales as close to possible, stock levels
retailers used to only rely on the previous years
The reason why predicting sales
past sales. This method is useful only to a certain appears to be so complex and
point and suffers many biases. Thankfully, Machine difficult is because, in a given period,
Learning has now evolved to be able to provide very many factors can affect purchase:
weather, purchasing trends,
accurate predictive models using different signals regulation, product launches, global
based on how they influence purchases. pandemic, buying behaviors ... And
the main issue with these types of
predictions based on past recordings
is that they don’t factor individual
incidents, and they make monthly
sales appear as if they were perfectly
distributed when they were probably
not. In fact, an out of order item might
have caused a slowing of sales of
a particular product or a particular
category, but it won’t show in the
monthly reports. Even worse, bad
numbers are viewed as a mark of
buyers disinterest, when they could
be the opposite as consumers over-
purchase an item and cause it to sell
out. It is also important to note that
a missing product in store doesn’t

35
necessarily mean that the product is
out of stock. Big box retailers struggle
to restock their shelves in real time
so a product that becomes instantly
popular might disappear from the
shelves and thus perform worse than
expected, when in fact, it is available
in inventory. Retailers are in need
of technology that can help them
step into a new paradigm that could
seamlessly align offer and demand.
Using Machine Learning to
help employees in stores

Machine Learning is the solution to

these challenges. Predictive models Concrete use cases
are now able to forecast sales up of Machine Learning on unavailable in shelves when they are
to four weeks by using a number of inventory management available in inventory. It is almost
signals that can affect sales, such as impossible for store managers to
the season, the current trends or the The technology is there. Now for ask their employees to monitor
price levels. The models are based on retailers to use it effectively and make shelves in real time and restock them
three indicators, the Day, the Product, accurate predictions, they need to be immediately. Technology solutions
and the Store. It is important that able to collect and analyse a huge using surveillance cameras and
they don’t rely on a single indicator amount of data. The problem lies weight sensors are a huge investment.
to be the most accurate possible. in the fact that they have different However, available information such
For the sake of the argument, sources of data and it can become as real-time sales, the characteristics
let’s imagine that you analyse the complex to try to process multiple of a SKU and the organization of a
season signal. Your data is going Excel and PDF files that contain particular shelf can be useful when
to be biased because predictions previous reports and media plans. predicting out of order. Models are
based on dates are not 100% helpful Retailers need to adopt big data tools able to analyze the usual flow of sales
because a certain date can fall on that can process this information of an item, meaning the time between
a weekday in a given year and fall into a clean and lisible digest, that two sales of the same product in a
on a weekend in the following year, is necessary to be able to create given store. In case of a statistical
and sales vary enormously between predictive models that can prevent anomaly, a human is notified for an
the weekday and the weekend. inventory issues. intervention in the store to analyze
Also, this date could end up on a the situation and solve it.
holiday, or a calendar event (such as The vast majority of retailers are
Christmas, Thanksgiving, Easter) or sitting on years of old sales data. There are many ways in which
a sports event. All factors that might However, this data can turn out to Machine Learning can help traditional
increase the consumption of some be inaccurate because of the effect retailers. Predictive analysis is only
products. As another example, let’s of past promotions or events (heat one of them. Retailers can gain a lot
have a look at the price levels signal. wave, sports event, major local event). from relying on advanced technology
Promotions at store level can severely To get rid of this bias, predictive to increase their store revenue by
affect the sales of a product from models are able to combine past better managing their inventory. But
a given category or even raise the sales numbers with those of similar processing vast amounts of data can
attractiveness of a store as a whole. stores. Artefact has enjoyed success also help them on the supply chain
Retailers need to be able to predict the with this approach and conducted a front, or on the merchandising side.
impact of their promotion strategy. successful experimentation in China Well-devised tools can become a real
That’s why it is important to take on the 020 platform that raised by asset to managers by taking care
into account many different signals 20% the level of accuracy in sales of complex and time-consuming
and indicators to accurately forecast predictions. tasks and provide them with accurate
sales through Machine Learning reports in a short amount of time and
and advanced Artificial Intelligence The other big challenge is to be letting them shift their priorities to
models. able to prevent items from being improve customer experience.

36
 DATA FOR RETAIL

Sales forecasting in retail:

what we learned from
the M5 competition.
In this article, Data Scientist Maxime Lutel sums
up his learnings from the M5 sales forecasting
competition, which consisted in predicting future
sales in several Walmart stores. He will walk you
through our solution and discuss what machine Maxime Lutel
learning model worked the best for this task. Data Scientist
ARTEFACT

Using machine learning in learning by doing, so we decided to categories. (cf Image 1)

to solve retailers’ business take a shot and code our own solution
Our task is to predict sales for all
to illustrate it. Now let’s go through
challenges the whole forecasting pipeline and
products in each store, on the days
right after the available dataset. It
Accurate sales forecasting is critical stop along the way to understand
means that 30 490 forecasts have to
for retail companies to produce the what worked and what failed.
be made for each day in the prediction
required quantity at the right time. But
even if avoiding waste and shortage Problem statement: horizon
is one of their main concerns, Hierarchical times series This hierarchy will guide our modeling
retailers still have a lot of room for forecasting choices, because interactions within
improvement. At least, that’s what product categories or stores contain
The dataset contains 5-year historical
people working at Walmart think, as very useful information for prediction
sales, from 2011 to 2016, for various
they launched an open data science purposes. Indeed, items in the same
products and stores. Some additional
challenge in March 2020— the M5 stores and categories might have
information is provided, such as sell
competition — to see how they could similar sales evolution, or on the
prices and calendar events. Data is
enhance their forecasting models. contrary they could cannibalize
hierarchically organized: stores are
each other. Therefore, we are going
The competition aimed at predicting divided into 3 states, and products
to describe each sample by features
future sales at the product level, based are grouped by categories and sub-
that capture these interactions,
on historical data. More than 5000
teams of data lovers and forecasting
experts have discussed for months
Image 1
about the methods, features and
models that would work best to
address this well-known machine
learning problem. These debates
highlighted some recurring issues
encountered in almost all forecasting
projects. And even more importantly,
they brought out a wide variety of
approaches to tackle them.
This article aspires to summarize
some key insights that emerged from
the challenge. At Artefact, we believe

37
and prioritize machine learning Image 2
based approaches over traditional
forecasting ones, to consider this
information when training the model.
Two main challenges:
intermittent values and an
extended prediction horizon
At this stage, you might think that
it is a really common forecasting
problem. You’re right and that’s why
it is interesting: it can relate to a wide
range of other projects, even if each
industry has its own characteristics.
However, this challenge has 2
important specificities that will make
the task more difficult than expected.
The first one is that the time series SEASONALITY
shifted by a certain period. For any
we are working with have a lot of Rather than using the sales date
specific item in a given store, the
intermittent values, i.e. long periods directly as a predictor, it is usually
1-week lag value would be the sales
of consecutive days with no sales, more relevant to decompose it into
made one week ago for this particular
as illustrated on the plot below. This several features to characterize
item and store. Different shift values
could be due to stock-outs or limited seasonality: year, month, week
can be considered, and the average
shelves’ area in stores. In any case, number, day of the week… The latter
of several lags is computed as well,
this complicates the task, since is particularly insightful because
to get more robust predictors. Lags
the error will skyrocket if sales are the problem has a strong weekly
can also be calculated on aggregated
predicted at a regular level while the periodicity: sales volumes are bigger
sales to capture more global trends,
product is out of shelves. (cf image 2) on the weekends, when people spend
for example at the store level or at
more time in supermarkets.
The second one comes from the the product category level.
task itself, and more precisely from Calendar events such as holidays
PRICING
the size of the prediction horizon. or NBA finals also have a strong
A product’s price can change from
Competitors are required to generate seasonal impact. One feature has
one store to another, and even
forecasts not only for the next week, been created for each event, with the
from one week to another within
but for a 4-week period. Would you following values:
the same store. These variations
rather rely on the weather forecast for • Negative values for the 15 days strongly influence sales and should
the next day or for 1 month from now? before the event (-15 to -1) therefore be described by some
The same goes for sales forecasting: • 0 on the D-day features. Rather than absolute prices,
an extended prediction horizon • Positive values for the 15 days relative price differences between
makes the problem more complex following the event (1 to 15) relevant products are more likely to
as uncertainty increases with time. • No value on periods more than 15 explain sales evolutions. That’s why
Feature engineering — days away from the event the following predictors have been
The idea is to model the seasonal computed:
Modeling sales’ driving impact not only on the D-day, but
factors • Relative difference between
also before and after. For example,
the current price of an item and
Now that we have understood the a product that will be offered a lot as
its historical average price, to
task at hand, we can start to compute a Christmas present will experience
highlight promotional offers’
features modeling all phenomenons a sales peak on the days before and
impact.
that might affect sales evolution. The a drop right after.
objective here is to describe each • Price relative difference with the
TRENDS
triplet Day x Product x Store by a set same item sold in other stores,
Recent trends also provide useful
of indicators that capture the effects to understand whether or not
information on future sales and are
of factors such as seasonality, trends the store has an attractive price.
modeled thanks to lag features. A
or pricing. lag is the value of the target variable • Price relative difference with

38
 DATA FOR RETAIL

other items sold in the same store

and same product category, to
capture some cannibalization
effects.
Categorical variables
encoding
Categorical variables such as the
state, the store, the product name or
its category also hold a significant
predictive power. This information

Image 4
has to be encoded into features to
help the model leveraging the dataset
hierarchy. One-hot encoding is not an
option here because some of these
categorical variables have a very high
cardinality (3049 distinct products). sales. Like many others, we have objective function is also available
Instead, we have used an ordered chosen another option, which is to for other gradient boosting models
target encoding, which means that rely on an objective function adapted such as XGBoost or CatBoost, so it’s
each observation is encoded by the to the problem: the tweedie loss. definitely worth trying if you’re dealing
average sales of past observations with intermittent values.
Without going into the mathematical
having the same categorical value. details, let’s try to understand why How to forecast 28 days in
The dataset is ordered by time for this loss function is appropriate for
this task to avoid data leakage.
advance?: Making the most
our problem, by comparing sales
distribution in the training data and
out of lag features
All categorical variables and some
of their combinations have been the tweedie distribution (cf image 3). As explained above, lag features are
encoded with this method. This sales shifted by a given period of time.
They look quite similar and both
results in very informative features, Thus, their values depend on where
have values concentrated around
the best one being the encoding of you stand in the forecasting horizon.
0. Setting the tweedie loss as an
product and store combination. If you The sales made on a particular day
objective function will basically force
wish to experiment other encoders, D can be considered as a 1-day lag
the model to maximize the likelihood
you can find a wide range of methods if you’re predicting one day ahead, or
of that distribution and thus predict
here. as a 28-day lag if you’re predicting 28
the right amount of 0s. Besides, this
days ahead. The following diagram
loss function comes with a parameter
Tweedie loss to handle illustrates this point (cf image 5)
— whose values are ranging from 1
intermittent values to 2 — that can be tuned to fit the This concept is important to
Different possible strategies can be distribution of the problem at hand understand what features will be
used to deal with the intermittent (cf image 4) available at prediction time. Here,
values issue. Some participants we are on day D and we would like to
Based on our dataset distribution,
decided to create 2 separate models: forecast sales for the next 28 days.
we can expect the optimal value to
one to predict whether or not the If we want to use the same model —
be between 1 and 1.5, but to be more
product will be available on a specific and thus the same features — to make
precise we will tune that parameter
day, and a second one to forecast predictions for the whole forecasting
later with cross-validation. This
horizon, we can only use lags that are
available to predict all days between
D+1 and D+28. This means that if we
use the 1-day lag feature to train the
model, that variable will also have to
be filled for predictions at D+2, D+3, …
and D+28, whereas it refers to dates
in the future.
Still, lags are probably the features
Image 3

with the biggest predictive power, so

39
 Image 5

it’s important to find a way to make

the most out of this information. We
have considered 3 options to get
around this problem, let’s see how
they performed.
OPTION 1:
ONE MODEL FOR ALL WEEKS
The first option is the most obvious
one. It consists in using the same
model to make predictions for all
weeks in the forecasting horizon.
As we just explained, it comes with
a huge constraint: only features
available for predicting at D+28 can
be used. Therefore, we have to get
rid of all the information given by the
27 most recent lags. It is a shame
as the most recent lags are also the
most informative ones, so we have
considered another option.
OPTION 2:
WEEKLY MODELS
This alternative consists in training
a different LightGBM model for each
week. On the diagram above, every
model is learning from the most
recent possible lags with respect
to the constraint imposed by its
prediction horizon. Following the MORE PRECISELY:
same logic as the previous option, it
means that each model can leverage • Model 1 makes forecasts for days 1–7, • Model 3 makes forecasts for days
all lags except those that are newer relying on all lags except the 6 most 15–21, relying on all lags except the
than the farthest day to predict. recent ones. 20 most recent ones.

This method allows us to better • Model 2 makes forecasts for days • Model 4 makes forecasts for days
capitalize on lag information for the 8–14, relying on all lags except the 13 22–28, relying on all lags except the 27
first 3 weeks and thus improved our most recent ones. most recent ones just like in option 1.

40
 DATA FOR RETAIL

solution’s forecast accuracy. It was

worth it because it was a Kaggle
competition, but for an industrialized
project, questions of complexity,
maintenance and interpretability
should also come into consideration.
Indeed, this option could be
computationally expensive and
if we are aiming at a rollout on a
whole country scale, it would require
maintaining hundreds of models in
live. In that case, it would be necessary
to evaluate if the performance
increment is large enough to justify
this more complex implementation.
OPTION 3:
RECURSIVE MODELING
The last option also uses weekly
models, but this time in a recursive
way. Recursive modeling means that simulate correctly out-of-sample training set and evaluated on the
predictions generated for a given accuracy, select hyper-parameters validation set using the root mean
week will be used as lag features for thoroughly and avoid over-fitting. squared error. (cf Image 9)
the following weeks. This happens When it comes to forecasting, this has
Folds 1, 2 and 3 aim at identifying
sequentially: we first make forecasts to be done carefully because there
parameters that would have
for the first week by using all lags is a temporal dependency between
maximized performance over
except the 6 most recent ones. observations that must be preserved.
recent periods, basically over the
Then, we predict week 2 by using In other words, we want to prevent the
last 3 months. The problem is that
our previous predictions as 1-week model from looking into the future
these 3 months might have different
lags, instead of excluding more lags when we train it.
specificities than the upcoming
like in option 2. By repeating the same
The validation period during which period that we are willing to forecast.
process, we always get recent lags
the model is tested also has a greater For example, let’s imagine that stores
available, even for weeks 3 and 4,
importance when dealing with time launched a huge promotional season
which allows us to leverage this
series. Model performance and the over the last few months, and that it
information to train the models.
optimal set of hyper-parameters can just stopped today.
This method is worth testing, but keep vary a lot depending on the period
These promotions would probably
in mind that it is quite unstable as over which the model is trained and
impact the model’s behavior, but it
errors spread from week to week. If tested. Therefore, our objective is to
would be risky to rely only on these
the first week model makes important find which parameters are the most
recent periods to tune it because this
errors, these errors will be taken as likely to maximize performance not
is not representative of what is going
the truth by the next model, which will over a random period, but over the
to happen next.
then inevitably be poorly performing, period that we want to forecast, i.e.
and so on. That’s why we decided to the next 4 weeks. To mitigate this risk, we have also
stick with option 2, that seems to be included folds 4 and 5, which
more reliable. Adapting the validation correspond to the forecasting period
process to the problem at respectively shifted by 1 and 2 years.
Ensuring model robustness hand These periods are likely to be similar
with an appropriate cross- because the problem has a strong
To achieve that goal, we have selected
validation: Why cross- yearly seasonality, which is often true
5 validation sets that were relevant
validation is critical for time to the prediction period. The diagram
in retail. In case we had a different
series periodicity, we could choose any
below shows how they are distributed
cross-validation strategy that has
In any machine learning project, over time. For each cross-validation
more business sense. In the end, we
adopting an appropriate cross- fold, the model is trained with various
have selected the hyper-parameters’
validation strategy is critical to combinations of parameters on the

41
 Image 9

combination with the lowest error incremental accuracy also depends engineering process: make sure
over the 5 folds to train the final on the order in which each step is that all the features you compute
model. implemented. will be available at prediction
time.
Results Key takeaways
• S elect a model architecture
The different techniques mentioned We have learned a lot from this that allows you to leverage
above allowed us to reach a 0.59 challenge thanks to participants’ lags as much as possible, but
weighted RMSSE — the metric used shared insights and we hope it gave also keep in mind complexity
on Kaggle — which is equivalent to you food for thoughts as well. Here considerations if you’re willing
a weighted forecast accuracy of are our key takeaways: to go to production.
82.8%. The chart below sums up the
• Work on a small but representative • Set-up a cross-validation strategy
incremental performance generated
subset of data to iterate quickly. adapted to your business problem
by each step.
•
B e super careful about to evaluate correctly your
These figures are indicative: the experiments’ performance.
data leakage in the feature

42
 WE OFFER END-TO-END
DATA & AI SERVICES

TRAVEL & TOURISM • FMCG • SPORTS & ENTERTAINMENT • BANKING & INSURANCE • RETAIL • LUXURY & COSMETICS
HEALTHCARE • eCOMMERCE • TELECOMMUNICATIONS • MANUFACTURING & UTILITIES • REAL ESTATE • PUBLIC & GOVERNMENT
CONTACT
[email protected]
artefact.com/contact-us

ARTEFACT HEADQUARTERS
19, rue Richer
75009 — Paris
France

artefact.com