Presentation on "A Complete Overview of Data Driven Decision Making in a Quickly Changing Business Environment" given by Isaac Aidoo, Head of Data Analytics, Zoona.
Report
Share
Report
Share
1 of 32
More Related Content
Day 1 (Lecture 2): Business Analytics
1. Business Analytics: A Complete Overview of Data Driven Decision
Making in a Quickly Changing Business
Speaker: Isaac Aidoo, Customer Analytics Manager (Customer Experience)
2. ● BA Mathematics & Statistics (University of Ghana)
● Head of Quantitative Research (MSR) - 1yr
● Retail Analyst & Advanced Analytics Lead (Nielsen)- 1.7yrs
● Business & Competitive Understanding Analyst (Tigo) -
11months
● Customer Analytics Manager(Tigo) - 1.3yrs
● Churn Manager(Tigo) - 9 months
● Manager, Data Services(Viotech) - 1.6yrs
● Manager, Customer Analytics(Consumer Experience) - 1 yr
PROFILE: ISAAC AIDOO
3. Focus areas
1. How companies are evolving - 10mins
a. Data
b. Personnel
c. Problems
2. Data Analytics Framework - 10 Mins
a. Big data definition
b. Data analytics types and trends
c. Data analytics framework
3. Case studies from Telco’s & Sponsored Data
a. Customer segmentation in Telcos - 20 mins
b. Churn prediction - 15 mins
4. Questions - 5 mins
4. How companies are evolving
● 2012 - Local Spirits Company Story
Previously:
Even in large organizations:
1. Traditional research was king
2. Excel was key
3. Schools were teaching traditional commercial tools like SPSS for
analysis
4. Typical size of data was not more than 65,536 rows over a 256 column
excel sheet
5. How companies are evolving
● 2015 - I was the 1st Customer Analytics Manager for Millicom Ghana
Now:
1. Competition has changed: Price convergence, Marginal product differentiation
2. Internet penetration - 33.6% from 2009’s 4% with over 4.9 million Facebook users
3. Huge data on customer calls: when, who and how often they call
4. POS systems capturing where and what customers purchase
5. Digital collection of KYC data & Social Media data
6. Now excel takes 1,048,576 rows over a 16,384 column excel sheet
7. R & Python are now thought in schools and a requirement for most Data analytics
positions
6. What we need now
1. C-Suite & Top Executive champions
2. Training
3. Networking
4. Experimenting
5. Showing business uplift
9. Definition
Business analytics (BA) refers to the skills,
technologies, practices for continuous iterative
exploration and investigation of past business
performance to gain insight and drive business
planning Beller, Michael J.; Alan Barnett (2009-
06-18). "Next Generation Business Analytics"
SKILLS
TECHNOLOGIES
DATA
ITERATIVE
PERFORMANCE
10. Levels of Analytics: Describe, Predict, Prescribe
Heroics ● Spreadsheets
● Extracts
Foundational
● Master Data Management
● Data Warehouses
● Data Governance
Competitive
● View Consolidated reports
● Dashboards
Differentiating
● Micro Segmentation
● Pattern Recognition
Break-away
● Mathematical optimization
● Reinforced learning
Descriptives
Customer 360 View
Predictive
(Predict the behaviour)
Prescriptive
Prescribe the
optimized action
Source: Analytics at work, Davenport et al
Analytics Maturity
Business
Impact
12. Moving from “Selling what we can” to “What they need”
Who am I?
What do I need?
When do I buy?
Where do I buy?
Demographics
Purchases
Community/Network
Interactions
Preferences
Intention to purchase
Purchase drivers
Purchase Triggers
Purchase Afinity
Activity Based
Life event based
Shopping Trip Type
Channels/Devices
Locations
Occasions
Who should I offer?
What should I offer?
When should I offer
How should I offer
Micro-Segmentation
and Personalization
Offer Allocation based
on Goal and
Constraints
Offer timing
Channel selection
CUSTOMER
NEEDS
ENTERPRISE
OBJECTIVES
Allocate
Optimized offer
15. CRISP-DM: Cross industry standard process
It provides a structured approach to planning and executing a
data mining/analytics project
We will use this to guide us through our case studies
17. Application areas in Telcos
1. Churn & Retention management
2. Product Development & Increase ARPU
3. Customer Acquisition
4. Customer Experience
5. Counter Fraud
18. Customer Segmentation
Business Problem: (Business objectives, Project Plan, Business success
criteria)
What are the customer product usage segments we have and does our
current product offering meet their needs. How can we develop the
identified segments to improve ARPU and customer retention
Business stakeholders: CVM and Product team, Marketing Team, DWH
team
19. Data Understanding
● Data on 3.9mil customers
● Data on Reload patterns, Products usage patterns (frequency &
amount), Community of Incoming and Outgoing Calls, Active and
Inactive days: A total of 43 variables
● Conducted initial descriptive analysis of all the different variables for
the 3.9mil
● Took a sample of 100k for the model largely because of resource
capacity
20. Data Preparation
● Majority of the behavioural data had considerably skewed distribution
so did some transformation of the data
● Took care of correlation in the data by conducting some principal
component analysis
21. Modeling
● In order to identify an optimal number of clusters to be found by the k–
means algorithm, several smaller random samples were first taken
from the training dataset and passed through the hierarchical cluster
analysis based on the Euclidean distance and using the Ward’s method
for agglomeration.
● This preliminary analysis showed initially that k=3 will give a more
stable solution (I however readjusted this to 5 after playing around
with the k and finding out that the 5 segments made business sense as
well)
● I then implemented k-means with which gave the 5 segments
23. Actions & Deployment
1. We commissioned a qualitative research to understand the
psychographics of these customers better and the reasons behind their
behaviour
2. The CVM team cleaned up all products and created new products
based on these segments
3. Liased with the DWH team to develop ways to implement this on the
larger customer base but this couldn’t materialize before I left the role
largely because we did not also have a CRM tool in place
4. This was later implemented in other African countries that Tigo was
operational
25. Customer Churn prediction
Business Problem:
The network was facing significant churn in customers and wanted to
identify:
1. Profile of customers who have churned
2. Based on that who is likely to churn
3. Identify if there are different segments within churners
4. Propose a product to win back churners
Business stakeholders: CVM and Product team, DWH team
26. Business Understanding ( Recall & Precision)
● Precision = How many of your positive
predictions are actually correct
Precision = True Positive / (True Positive + False
Positive) : When the cost of predicting that
someone will churn is high then you might want a
better precision
● Recall = How many of the positive cases did you
get correct.
Recall = True Positive / ( True Positive + False
Negative) : When the cost of predicting that
someone will not churn is high
● Without modeling because of the
hugely imbalance nature of the data
(4% of customers were churners) not
building a model but randomly
picking customers would have
yielded 96% accuracy. So we had to
look at different metrics
27. Business Understanding
● After engaging stakeholders, business wanted a higher recall measure
for HVC customers because of the cost associated with (False
Negatives) - When the model predicts that customer will not churn but
actually churns; For other segments MVC/LVC we agreed on higher
precision as company wasn’t ready to give freebies enmass to
customers who don’t give much and
28. Data Understanding
● Data on 179k churners over 4.5mil customers
● Data on Reload patterns, Products usage patterns (frequency &
amount, time of call, duration), VAS, Community of Incoming and
Outgoing Calls, Active and Inactive days, 59 variables
● Hold out recent two months data as test set
29. Data Preparation
● Most customers do not suddenly stop using the service we had to break
transactional data into periods and create customer cadence metrics for
most of the transactions data
● Majority of the behavioural data had considerably skewed distribution so
did some transformation of the data
● Preliminary feature importance
30. Modeling
● Different models were built for HVC customers and (MVC & LVC
customers)
● C5.0 Decision tree model was used, initially to understand the path of
the churner
● Logistic regression, SVM, Adaboost were the models implemented
● SVM gave the highest precision on the test set at 89% but recall of 68%
● C5.0 gave a recall of 89% and Precision of 70%
● We also realized that the community of customers was a major factor
in predicting churn
31. Actions & Deployment
1. Created a new product where one onnet contact of customers who had
high likelihood of churning were given free airtime to call and activate
the churner
2. Created a churn and customer retention framework with relevant
product offerings