Using image recognition algorithms, open data extracted from internet sources and social networks, and pricing algorithms we helped clients to develop a marketing and pricing strategy to improve their brand image and increase sales in the HORECA channel.
In the company where I work, Minsait, we develop business intelligence solutions and we go hand by hand with our clients to help them in their digital transformation strategies, route to market plans and pricing strategies. In this talk I want to develop one concept that we developed for a customer that is one major Spanish company working in the FMCG field.
Success case of modern algorithms applied in FMCG industry - Alberto Molinero Zaera
1. 1
How data science can solve real
life business case scenarios
Success case of modern algorithms applied to a particular
business case in the FMCG industry.
September 2018
2. 2
Who am I?
Alberto Molinero Zaera
Studies
⢠Aerospace Engineering (Universidad PolitÊcnica de Madrid, UPM)
1. MsC Spacecraft and Missiles (Universidad PolitĂŠcnica de Madrid, UPM)
⢠Master in Business Project MGMT (ESCP Europe)
⢠Mecenos
⢠SAP Young profesional graduate program
⢠Minsait Strategy consulting analyst
4. 4
The main focus of the Project was the HoReCa (Hotels-Restaurants-CafĂŠs) channel
where the client was unable to adapt himself to the quick changes that happened in the
last years
What?
⢠New small players selling
Good quality products at
lower prices
⢠New products and ranges
(eco-friendly, premiumâŚ)
⢠Regional players turning
national
How?
⢠Innovation as a market driver
⢠Need to create new products
to satisfy new needs and
clients
⢠Complex market
segmentation
⢠Social media (influencers,
FacebookâŚ) as driver for the
Brand awareness
Why?
⢠Weak Brand Image
⢠Aggressive marketing
strategies by competitors
⢠Lack of resources to turnaround
the situation (reducing prices)
Really?
⢠Primitive market
segmentation
⢠Lack of data driven KPIs to
measure the impact of the
different campaigns
⢠Unoptimized pricing
startegies
Traditional
Strategy
Consulting
Data Science
SUCCESS
Our client is a big player in the food & beverages spanish industry facing several problems
Lack of BI tools
and KPIs
Changin market: more
dynamic and complex
Lost of clientsIncreasing
competition
6. 6
Machine learning algorithms where used to create a customer segmentation, pricing
models and an optimal product mix for every client in the spanish HoReCa channel
More details
below
⢠Assestment of the current state of the clientâs data, quality and mining
⢠Scrapping bots to collect information from open sources (Social Networks,
gastro guides
⢠Third party data (Nielsen, Spanish statistics institute, external providersâŚ)
⢠Creation and implementation of Extract, Transform and Load processes
⢠Segmentation and clustering
⢠Price elasticity modelling (pricing)
⢠Optimal product introduction
⢠Client segmentation in the HoReCa channel using convolutional neural
networks and images obtained from social media profiles
Dataset creation
ETL processes
Machine
learning
Image
recognition/Deep
learnign
7. 7
After applying machine learning algorithms we developed a more complex market segmentation
suitable to create customized marketing strategies and policies adapted for each client
1. Traditional restaurant
2. Bar restaurant
3. Gastro bar
4. Haute cuisine
5. Modern restaurant
6. Tapas bar
Traditional approaches such as random forests, decission tres and regressions can be implemented
easily and fast with reasonably good results. One solution can be formed by a combination of boosting
algorithms, regressions and neural networks.
A classical usupervised learning algorithm such as K-Means or SVM not always gives the best approach in a
business case scenario. Instead we analyzed the market and defined 18 customer segments with which we
created a first data set to train unsuppervised learning method.
The output was a normalized score showing the probability of each client of being one of the segments defined.
Market
segmentation
Business oriented
segments: supervised vs
unsupervised learning
18 segments to
define the spanish
HoReCa universe
Algorithms
Scoring
1. Ethnic
2. Fast food
3. Brewery
4. CafĂŠ
5. Traditional bar
6. Sports bar
1. Afterwork
2. Early Night
3. Night Club
4. Tavern
5. Microbrewery
6. Bar tradicional
8. 8
The delivearable was a complete database showing a segmentation profile for more tan
100K Points of sale (POS) in the spanish market. This allow the client to define micro
marketing strategies and optimum product introductions in each POS.
Tatel- Madrid Txakolina- Madrid Modern restaurant
Haute cuisine
Bar restaurant
Gastro bar
Traditional restaurant
CafĂŠ
Tapas bar
Traditional bar
Fast food
Nightclub
Afterwork
Early night
Sports bar
Microbrewery
Brewery
Ethnic
Tavern
Haute cuisine
Bar restaurant
Gastro bar
Traditional restaurant
CafĂŠ
Tapas bar
Traditional bar
Fast food
Nightclub
Afterwork
Early night
Sports bar
Microbrewery
Brewery
Ethnic
Tavern
Modern restaurant
9. 9
Price elasticity modelling is a complex data science problem that requires precise data
and an extensive database to be able to model all the parameters that affect the duo
Price-demand
It is necessary first to asses the database of the client to ensure that
the quality of the data is good enough to train such a model:
⢠Competitors prices
⢠Transaction volumes
⢠Retail prices
Data maturity
What a pricing tool is
And what is not
Price elasticity
modelling
⢠Costs
⢠Promotions and activations
Real Price (test)
Prediction
I. Optimizer of prices and pricingâs team workload
II. Compliance with pricing policies
III. Price differentiation at deal level.
IV. Endless Price monitoring/reporting.
Pricing results obtained using
XGradient boosting algorithms
I. Risk of low acceptance due to complexity (black box problem)
II. Lack of automatic implementation of pricing strategies
III. Business mentality analysis is necessary
IV. Restricted capabilities to reflect market changes and customerâs willingness to
pay
11. 11
The current market segmentations requires constant updating of the databases with new
data about new POS. This is a slow and expensive process. Computer visiĂłn could help to
reduce costs without compromissing the quality of the data obtained
The Google Vision API delivers a
fast and cheap first approach to
obtain some basic labelling in
images extracted from Social Media
Google Vision
Customized
modelling
Social Network images
(Train set)
Tensorflow allows an easy and
quick implementation of
Convolutional Deep neural
networks to crĂŠate customized
labels
Market segmentation replica
using the predicted labels