This is the National wildcard winning solution for Airtel iCreate competition. We were selected amongst Top 5000 teams participating from all B-schools over India. We presented the only analytical segmentation driven customer strengthening model.
Our proposal was to use Gradient boosting to identify churn, positive revenue customers to turnaround the customer strategy.
We wanted to drop the low value customers and focus on increased services to the high value customers by offering value driven offerings.
Heart Disease Classification Report: A Data Analysis Project
Airtel iCreate National Wildcard Winners 2019
1. airtel iCreate 2018
Wild Card Case Study : Strengthening the Customer Base
(Finance)
Team Name: StormTroopers
Campus: IIM Rohtak
Members: Naveen Kumar, Saurav Sarkar
2. 1. Poor Customer Additions
Financial & Non-Financial Metrics
2
Fintel has seen Competitive boundaries shifting as it’s core businesses comes under intense competition due to the emergence of
a new competitive player in the market & regulatory pressures such as MNP resulting in low switching costs for users. This
resulted in Poor Customer additions with low value per user affecting Fintel’s Financial & Non-Financial Metrics
Increasing CapEx puts
redundant pressure to
satisfy increasing data
needs
Financial Terms Implications on Fintel Non Financial Terms Implications on Fintel
LeadingIndicatorsofFuturePerformance
Subscriber Acquisition Cost : Dealer
Commission, Terminal Subsidy, Sales &
Distribution
This is the direct measure of the Fintel’s expenditures on acquiring new
customers through it’s operators trade channels. Due to the arbitrage
opportunities present in the Customer hopping between multiple network
operators, this GAC cost has risen in the last few quarters
101 84 87 80
0
100
200
Q4
GAC
Q3 Q2 Q1 Operating Profit margin is taken a direct
hit due to unstructured payouts affecting
Fintel’s retention efforts & increasing CAC
Fintel’s Business strategy of giving
high offers to Low Customer life
time value segment with high
switching periods directly affects
EBITDA Margins
Low Capital Productivity
Low ROIC (EBIDTA/ Gross Capex)
Decreased Gearing (Debt/ EBIDTA)
Take
Rate
Adoption
rate
Take Rate signifies how many customers/prospects act on the offers
rolled out, sign up for new products & Innovation is the ability to bring new
products/services to market successfully -> Both these metrics reflect
the company’s ability to bring value to your customers & the market.
Revenue
Diversification
“Foot in the door” Channel Expansion
Revenue Synergies
Subscriber growth is one
of the main Non Financial
Metrics affected by the poor
customer additions which
directly affects the
operating margins and the
portfolio of services that
Fintel has in it’s Pipeline
threating sustainability
New Revenue models: for future rewards
• High leveraged capital structure
• Focus more on OPEX & CapEx uses
Measure of customer satisfaction across the
Organization
Decreasing focus on
improving customer
satisfaction due to
latency affects the High
CLV Segment
Subs Segmentation: Postpaid to Prepaid switching is increased due to
switching of price sensitive -> Increased churn (Dual Sim Mobile Factor)
No longer makes sense to invest as heavily as
once made sense, since revenue gains will
underperform the intensity of the investment.
That is a strategic issue of paramount importance
to Fintel in face of rising data & voice usage by
Customers
Customer Switching hits
the overall strategy which
affects future growth
The Level of new product
rollouts which Fintel can
bring to the market is
affected due to low
reserves
Drop in the redirection time
and CLV usage of
customers resulting in high
Churn Rates of Fintel
Share of revenue declining
due to the rotational sim
resulting in bottom line
getting affected
Increased spending and
decreasing productivity
threatens long term
sustainability encumbered
by debt costs reducing
expansion plans
Source: https://www.ey.com/Publication/vwLUAssets/EY-telecom-analytics-advisory/%24File/EY-telecom-analytics.pdf
https://smallbusiness.chron.com/adjusting-paper-margins-powerpoint-29281.html
3. 2. Trade Channel Fraud Analytics
Management System Solutions
3
Fintel’s Incentive payout schemes has been taken advantage of by players & the problem is widespread throughout the Trade
Channel resulting in increased Customer Acquisition costs for Fintel slashing it’s profit margins. The commingled effect of
rotational sims give rise to a significant arbitrage opportunities for traders creating a vital need for a targeted action program
Intensive Growth Strategy
Customer Integration
Trade Channel Diversification
Base Case
Strategy Planning Gap
Targeted Action
plan to ensure
accountability
is in absentia!
Increase Customer
acquisition
Decrease subscribers
acquisition costs (SAC)
Arrest Churn rates
Trade Channel
Fraud Tactics Study
1. Distributors and POS
activating SIMs
previously to selling
them through non-
regular distribution
Networks at a lower
price
2. Fake activations and
signing fake contracts
3. Distributors making
Offnet calls to the
previously activated sim
to get the consumption
commission on top the
activation and contract
fulfillment ones
Point of establishment at 3 check points on the sales & distribution chain
1. SIMs delivery by the Logistic Operator to the Distributors
2. Activation
3. RGS (register of revenue generation call) & Receipt of the signed contract
Wallet Size
Strategy
A predictive analytics model can be built using network
event data (mediation devices), billing data (billing system),
customer data (CRM) and payments data (accounts
receivable) to predict the likelihood of potential fraud in the
future.
Source: https://pdfs.semanticscholar.org/33c6/186be37dfd1422c04dedb8b8f20bede95d33.pdf
http://www.flytxt.com/blog/predictive-customer-churn-modelling-in-telecom-industry-with-greater-accuracy/
https://www.urbanairship.com/blog/churn-prediction-our-machine-learning-model
4. 4
Fintel’s Analytical Capabilities are not robust enough to support effective monitoring in it’s trade channel. We have proposed an
integrated Analytics solutions which Fintel must incorporate into it’s strategy to maximize it’s throughput. We recommend
sophisticated fraud scoring engine applied with risk- and value-based scoring models to bolster sophisticated challenges
2 (Cont.) Analytical Management
Solutions and related KPIs
Channel Fraud
KPI’s for Trade Channel monitoring
Activations per POS /
Distributor
RGS =
Revenue
Generating
g
Subscriber
RGS per POS /
Distributor
Activation and RGS
location (city)
RGS generated with
incoming/outgoing
calls
Contract validity
Telecom Policy and
high license fees
-
-
-
-
-
-
Uncertainty
regulatory policy
Worsening CapEx
to support growth
Access to advanced
analytical capabilites
Competition
from new
technologies
Cost
containment
Customer
shifts
Human
Capital
deficitNew
Operational
challenges
Price
Volatility due
to
competition
-More Same Less
P
R
E
D
I
C
T
E
D
R
I
S
K
L
E
V
E
L
Increase collaboration between
the fraud, marketing and
finance/credit risk
departments.
Through efficient customer
screening and scoring, all internal
teams affected by fraud can start
working together to solve these
problems – with minimal
interference to the customer
journey. Analytical models can
help prevent the use of high-
end devices to fraudsters who
should not become customers.
They can help identify different
offerings to different customers
depending on their risk profile
A Base case example of using Analytical
capabilities to curb Multi-Sim Rotational
Problem of Fintel
These KPI’s serve as control points
and when introduced on the chain will
have full control over the chain’s
activity.
They would help monitor the
implementation process, providing
Fintel relevant feedback to fine tune
the different actions to ensure
efficiency.
An Encompassing Analytical Customer
quality model framework
Leveraging Analytics to curb user switching
Source: R. Fildes, “Telecommunications Demand Forecasting - A Review,” International Journal of Forecasting, vol. 18, 2002, pp. 489-522.
Team StormTroopers Student Research
5. Customer segmentation
5
Segmentation based on customer value journey (LTV Model)
Acquisition Growth RetentionOnboarding
Efficient
Acquisition
Positive
Engagement
No
Response
Failed Costly
reengagement efforts
Continued engagement
through VIP treatment
Good
quality
custom
ers with
High-
CLV
Asking for more VIP
plans with offers
Mid-
Level
CLV
Custom
er
Low-
CLV
custom
ers
RFM Model for Customer segmentation
Customer Value Pyramid
Top
Big
Medium
Small
Inactive
Highly
Active
Medium
Active
R:
Regency
F:
Frequency
M:
Monetary
Hybrid Model for Segmenting customers
Custome
r Profile
RFM
Model
Generic
Algorithm
Selection
Crossover
Mutation
Fitness
Value
Customer
Segmentation
Strategic Design for
segments
Strategic
Execution
LTV
Model
Regency, Frequency and Monetary (RFM) model and lifetime
value (LTV) model to evaluate proposed segmented customers.
For selecting more appropriate customers for each campaign
strategy, this work proposed using generic algorithm (GA)
Behavioral Segmentation( 1st step to identify Customers)
Plain Loyal:: A
customer that has
always been Active
Not Dependable:
reached the Churn
status for the first time
Fence Seated: Moved
out of Active into
Dormancy for first time
Loyal Under Incentives:
Became active again due
to offers
Due to inconsistent behavior of consumers, we need to look into Customer Lifetime Value
We can easily segmented customers through their behavior. After that we can get the Life Time Value of each segment. Through this CLV
model and RFM( Revenue, Frequency & Monetary) model we designed a model which will work on Generic Algorithm to segment the
customers properly. After that we can sub-segment those based on their data usage basis. Thus we can get the proper view of segmentation
Potential Customer Segmentation based on
internal CLV Usage of Data Traffic Model
Source: https://www.researchgate.net/publication/262056572_A_Customer_Churn_Prediction_Model_in_Telecom_Industry_Using_Boosting.
Team StormTroopers Student Research
6. Predictive Churn Model
6
Historical
Data
New Data
Test Set
Model
Training/
BuildingTraining
Set
Test
Model
Prediction
Deployed
Model
CLV
Prediction for
next Quarter
Customer Data
CRM
Call Centre
Records
Application
Logs
Web
Clickstream
Data
Discovery
Model
Production
Model
Customer Lifetime Value Prediction for next Quarter
Give rating
according
to CLV
CA BA AA
CB BB AB
CC BC AC
Customer Segmentation through rating
Stickiness Index
Revenue per
Customer
HCV
Customer
s
LCV
Custo
mers
Interdependent Data for Estimating Churn Probability
Customer
HLV MLV SLV
ACCA BA AB
BCCC CB BB
Churn Rate Calculation=
Average Churn Rate for the segment in previous quarter*Average CLV for the segment
Decision Tree for Predicting the Customer Churn
Predictive
Positive(c=1)
True
Positive(TP)
False
Positive(TP)
Predictive
Negative(c=0)
False
Negative(FN)
True
Negative(FN)
Actual
Positive(y=1)
Actual
Negative(y=1)
Accuracy = (TP+TN)/(TP+TN+FP+FN)
Recall = TP/(TP+FN)
Precision= TP/(TP+FP)
F1 score= 2*(Precision-
Recall)/(Precision+ Recall)
Ca: Cost of contacting the customer
Co: The average cost of the retention offer
CLV: Customer Lifestyle Value (Considered as intangible Asset)
New
Customer Active
Customers
TP
FP
FN
TN
Effective
Customer
C0+Ca
C0+Ca
0
CLV+Ca
CLV
Inflow
Customer
Base
Predictive Churners
Predictive Non-Churners
Outflow
Resource
Lost
Cost of Acquisition < Cost of
Retention
Increasing Churn
Rate
More spending on
Customer Retention
High Variable
Cost
Discount offering for
satisfying customers
High Fixed cost
due to high
Capex
Difficult to reach
Break-even point
ARPU & ARMU goes down
Operating Free Cash
Flow declines
EBITDA Margin
comes down
Profit will
decline
Non Controlling
Interest increases
Acquire
customer
from
Competitor
High Fixed
Asset
High cost
for upselling
to new
customer
High Marketing
Cost
High Maintenance
Cost
High Operating
Cost
Loss of
Recurring
Revenue
Expense Increasing
Voice Usage Call
Recharge
Pattern
Data and
VAS
usage
SMS Usage
Location &
Network
Age
Churn Rate affects the Bottom-Line Financially
Exploratory data analysis (EDA)
For predicting customer churn with previous quarter data we divide the customer in segments, then make sub segment of those. Then
through decision tree & supervise machine learning we will predict the churn rate of each sub segment. We are giving grade to customers
based on Revenue per Customer & Stickiness Index of Customers, A is the highest & C is the lowest
EBITDA =(Revenue)-(cost of sales)-(service costs)-(network costs)-(bad debt expenses)-(marketing costs)
Financial Impact of Cost of Churn
Source: https://www.tcil-india.com/public/pdf/AR-2016-17.pdf
Team StormTroopers Student Research
7. Exhibits
7
Prevention
& detection
Tools
Data
Visualiz
ation Behavio
ral
Analysis
Deep
Learning
Forensic
Tools
Investig
ation
CellsFlexible
Audit
Plans
Internal
Controls
Bench
marking
Real
Time
Screeni
ng
Automat
ed
Controls
Complia
nce
Solution
s
Third
Party
Screeni
ng
Fraud
risk
Assess
ment
Vigil
mechanis
m
Awarenes
s
Initiatives
Governan
ce
EXHIBIT 1: Important fraud prevention and detection tools
1. Integration, quality and management of data. CSPs are facing a high volume of data internally (CDR, customer data,
etc.). But as noted above, they can also use external data, including both structured (lists, demography statistics) and
unstructured (social media information, web crawls) data. Combining these data sources into a single system for
analysis requires a system that can access data sources transparently and handle the data quality tasks to make sure
that the data is high-quality and ready for analytics.
2. Detection. Using a hybrid analytical approach that combines business rules, anomaly detection, predictive
modeling and social network analysis, organizations can quickly detect various types of fraud. By helping the
system “learn” from the results of each round of analytics
3. Alerts qualification and investigation. Alerts should be clear and understandable, giving fraud investigators the
opportunity to visualize and prioritize the alerts. For each of the alerts, the investigator needs to clearly understand
the reasons of the alerts to shorten the investigation time and process. After investigation, the outcome will be sent
back to the system to increase the detection logic –also known as the feedback loop.
4. Monitoring and ad hoc analysis. It is critical to measure program performance and monitor key performance
indicators of successful fraud prevention. Also, following up on existing fraud instances can be used by
management to determine future fraud prevention approaches. Monitoring the program over time and updating
fraud prevention models will also help formulate long-term strategies.
5. High-performance and near-real-time technology. The underpinning of the Fraud Framework is its ability to run
analytics on a very large number of transactions consistent with the communications industry and do this in near-real
time (intra-day)
Source: Team StormTroopers Student Research
8. Exhibit
8
Target
Customer
New
Customer
Initial
Customer
High Value
High
Potential
Low Value
Voluntary
Churn
Forced
Churn
Acquisition Activation Relationship
Prospect New Customer Established
customer
EXHIBIT 2: Customer Life-Cycle in Telecom Industry
Win Situation
for Company’s
Mobile
High Tech
Price
sensitive
Foreigners
Low Tech
Loyal
Analogue
Treacherous
Analogue
Prevention Retention
Latest technology Offers ( broadband,
gadgets etc.)
• Cross Selling
• Up-Selling
• Weekend, evening calls to bind
• Discounts to push to high-tech segment
Invest less to keep
CustomerValue
Propensity to churn
Identify customer behavior
Identify propensity to churn
Identify customer profitability
EXHIBIT 3: The customer portfolio based on customer value,
propensity to churn and identified clusters
Network Operation
Business
Support
Function
Collaboration
across value
chain
Sales,
distribution,
and customer
service
IT
Infrastructure
Products, pricing, and marketing &
End-to-end redesign
IT Services, Billing ,
Customer Care
Joint venture on
marketing and
media
Digital distribution and
customer service
Outsourced
customer service Joint procurement
venture
Managed services
partnerships
Managed services for
finance operations
Passive
network
sharing
Active network
sharing
Operator
• Higher share of
variable cost
• More flexible to
fluctuating
volumes
• Cost and asset
base become more
flexible, enabling a
better competitive
stance outside of
strategic
capabilities
• Value chain is
significantly
transformed, from
mostly internal
delivery into a
flexible modular
set-up
EXHIBIT 4: How Cost of Churn Affects the Telecom Operators
Customer data
Offer acceptance
prediction
Collection risk
estimate
Revenue
estimate
Churn prediction
Decision making
Decision making
Churn
intervention
Profitability
optimization
EXHIBIT 5: THE FRAMEWORK FOR CUSTOMER RETENTION AND
PROFITABILITY MAXIMIZATION
Source: Team StormTroopers Student Research