2. Work Abstract
www.valiancesolutions.com
Product recommendation
model for prominent Life
Insurer identifying top 2
products existing customers
are likely to purchase.
Analysis was used in email
and direct marketing
campaigns.
Prediction Model for
identifying customers who
are unlikely to pay insurance
premium within 30 days
grace period. Results were
used to formulate pro-active
customer retention
strategies.
Monthly sales forecasting
model for prominent direct
sales retailer in US using
Neural Networks. Achieved
average forecasting
accuracy of 7 percent with 5
to 10 percent error range.
Cross Sell Customer Churn Sales ForecastFraud Prevention
Real time Fraud Detection
algorithm for unsecured
consumer lending.
Substantial decrease in loan
disbursement to fraudulent
cases at Point of Sale
3. Case Study: Product Recommendation
www.valiancesolutions.com
Company Profile
Project Profile
Technology Used
Business Need Solution Benefits to Customer
Location
India
Industry
Insurance
Project type
Propensity Modeling
R
SQL
Excel
To identify the propensity to cross-
sell a policy
• To proactively identify the policy
holders who have high likelihood
to purchase more than one policy
• Use agent characteristics as main
lever to predict cross-sell
propensity
Propensity Algorithm to score
customers using Logistic
Regression
• Cross Sell propensity scores at
product category level for each
customer.
• Scores normalized to recommend
top 2 products customer is likely to
purhacase.
• Recommendations used to power
email and call center campaigns.
Tailored marketing campaigns
across modes of marketing
• Efficient Marketing Campaigns
• Incremental Revenue of USD
100,000 in 3 months
• Lower cost of Marketing
Campaigns
4. Case Study Details
www.valiancesolutions.com
Input
Data
Data
Cleaning
Exploratory
Data Analysis
Data
Enrichment
Propensity
Modeling
Algorithm
Implementation
Customer Attributes
Product Attributes
Transactional
Behavior
Interaction behavior
Missing value
Treatment
Correcting incorrect
values
Removal of duplicate
records
Uni-Variate Analysis
Bi-Variate Analysis
Creation of new
variables
Variable
transformations
Multiple versions of
Models basis
different variable
selection
Model Comparison
Choice of best model
Modify marketing
campaigns.
Feedback monitoring
Algorithm tweaking
(if needed)
Solution: Propensity Algorithm to score customers using Logistic Regression
Objective: To identify the propensity to cross-sell a policy
To proactively identify the policy holders who have
high likelihood to purchase more than one policy
Use agent characteristics as main lever to predict
cross-sell propensity
5. Cross Sell Model
www.valiancesolutions.com
Illustrative
All the customers
acquired in
Analysis Window
Characteristics
Characteristics
Scoring model
Likelihood to Cross-sell
Scoring Algorithm for
Calculation Propensity to
Cross-sell
Identify the Last Agent of a particular customer for Agencies- which
maximize propensity to cross-sell
Customers holding multiple policies in
Analysis window
Customers holding single policy in
Analysis window
If a customer cross-sold more than one policies during analysis window, then each
cross-sell instance will be considered as cross-sell opportunity (one customer might
appear more than once in modeling window)
Identify Best Agent for ARD - which maximize propensity to cross-sell
Orphan Customers of Agencies*
Cross-sell
Campaigning
7. Cross Sell Solution
High Purchase Propensity
Medium Purchase Propensity
Low Purchase Propensity
Tailored marketing
campaigns across
modes of marketing
Efficient Marketing
Campaigns
Incremental Revenue
of USD 100,000 in 3
months
Lower cost of
Marketing Campaigns
Cross Sell
Algorithm
www.valiancesolutions.com
8. Case Study : Customer Retention
www.valiancesolutions.com
Company Profile
Project Profile
Technology Used
Business Need Solution Benefits to Customer
Location
India
Industry
Insurance
Project type
Lapse Modeling
R
SQL
Excel
Logistic Regression
Random Forest
Improving Customer Retention
• To identify policy holders who are
likely to lapse and move out of the
program
• Take proactive measures to keep
them in the program
Quantitative Analysis of Lapsation
• What are the reasons for attrition?
• What are patterns in customer
attrition across different tenure of
policy?
• How does the attrition rates change
by changing factors?
• What is the probability of a customer
to attrite?
• What channel or combination of
channels which will deliver the most
conversion?
Churn Scoring algorithm based on
machine learning.
• Upcoming renewals scored on
monthly basis in a batch mode.
• Customer Segments created on
basis of churn score and Annual
Premium.
• Contact Strategy finalized on basis
on churn score and premium at
stake.
• Customers with higher churn
score and premium >25k pursued
through calls and visits if needed.
• Customers with lower churn score
and lower premium contacted via
sms and emails.
• Frequency of emails, call s to be
adjusted as per segment.
Customer Churn
• Policy Persistency increased by
20% over 1 year
• Incremental Revenue of 3M USD
in 1 year
• Lower cost of retention
Campaigns
9. Case Study Details
www.valiancesolutions.com
What are the
reasons for
attrition?
What are patterns
in customer
attrition across
different tenure of
policy?
How does the
attrition rates
change by
changing factors?
What is the
probability of a
customer to
attrite?
What channel or
combination of
channels which
will deliver the
most conversion?
Quantitative Analysis of Lapsation
Objective : Improving Customer Retention
To identify policy holders who are likely to lapse and
move out of the program
Take proactive measures to keep them in the
program
10. Solution: Lapse Model
www.valiancesolutions.com
Illustrative
Policies for
renewal between
Analysis Window
Characteristics
Characteristics
Scoring model
Likelihood to lapse
Policies lapsed between Analysis
window are bad
Policies lapsed between Analysis
window are good
Retention
Campaigning
Application on policies
coming for renewals in
following month
Scoring Algorithm for
Calculation Propensity
to lapse
Lapsed and Reinstated
Lapsed
Non Lapsed
11. Sample Deliverable: Customer Risk Profiling
www.valiancesolutions.com
Illustrative
Customers were segmented on basis the probability to lapse and APE band
APE BAND
Risk Group <18K
Between 18K
and 25K
>25K Total
High 18% 8% 14% 40%
Medium 15% 8% 7% 30%
Low 10% 7% 13% 30%
Total 43% 23% 34% 100%
Customers were segmented in High,
Medium and Low risk profiles on basis
of Annual Premium and their
probability to lapse.
Cut off probability band for High,
Medium & Low group was identified
from customer deciles. i.e. For High
band probability cut off was based on
top 30 percent of lapsers.
Proactive campaigning to customers
with higher likelihood to lapse
Risk_Group Probability of Lapsation
H >0.18
M 0.03-0.18
L <=0.03
High Risk Priority 1
Medium Risk Priority 2
Low Risk Priority 3
Legend
12. Customer Churn Solution
High Churn Propensity
Medium Churn Propensity
Low Churn Propensity
High risk customers to be
reached pro-actively through
calls and visits if needed.
Medium risk customers to be
reached through calls, emails
and sms’s
Low risk customers to be
reached through sms’s and
emails.
Policy Persistency
increased by 20%
over 1 year
Incremental Revenue
of 3M USD in 1 year
Lower cost of
retention Campaigns
Churn Propensity
Algorithm
www.valiancesolutions.com
13. Case Study : Fraud Modeling for Unsecured Loans
www.valiancesolutions.com
Company Profile
Project Profile
Technology Used
Business Need Solution Benefits to Customer
Develop Credit Risk framework for
POS loan approvals
• To identify customers who are
more likely to commit fraud/default
on consumer durable loans.
• To streamline loan approval
process according to customer
risk profiles.
Real time Fraud Propensity score
at Point of Sale
• Machine Learning based fraud
engine integrated with CRM
• Assigns fraud score for applicant
at point of lending.
• Higher fraud score applications
routed through stringent
verification process.
Substantial decrease in fraud thus
improving the Bottom Line
• Substantial decrease in loan
disbursement to fraudulent cases
at Point of Sale
• Almost 10% of the originations are
referred to ‘Normal process’ in
which the fraud incidence is as
high as 5% which translates into a
gross saving of almost 1.5 million
USD i.e. 50% of the VaR
• Substantial decrease in the third
party cost of loan amount recovery
from the fraudulent cases.
Location
India
Industry
Banking
Project type
Fraud Likelihood Model
SAS
SQL
Java
Excel
14. Case Study Details
www.valiancesolutions.com
Identify attributes of customers who
are most likely to commit fraud?
What are patterns in customer
default across cities/income/
profession segments?
What is the probability of a
customer to default?
Quantitative Analysis of Credit Risk
Objective: Develop Credit Risk framework for POS loan approvals
To identify customers who are more likely to commit
fraud on consumer durable loans.
To streamline loan approval process according to
customer risk profiles.
15. Solution
www.valiancesolutions.com
Text Mining
Fraud Likelihood
Model
Development of
technology solution
Implementation
framework
Strategy roll-out
and testing
• Hypothesis building
• Data cleansing
• Conducting field visits
to understand typical
trends in fraud
patterns
• Profiling patterns
• Algorithm for fraud
prediction
• Build a Java based
algorithm
• Ensure compatibility
with client’s Sales
CRM system
• Host the algorithm on
the client’s system
• Cross-validate the
scores generated by
the system
• Roll-out the algorithm
on the live system
• Continuous monitoring
of through the door
population for any
changes in patterns
16. Fraud Likelihood Model
www.valiancesolutions.com
Illustrative
All account
sourced
Characteristics
Characteristics
Scoring model
Likelihood to Default
Customers identified as not fraud
Customer s identified as Fraud
Loan application coming
for renewal at POS
Scoring Algorithm for
Calculation Propensity
to default
Medium Risk
High Risk
Low Risk
17. Implementation Framework
www.valiancesolutions.com
Customer walks-in to outlet
for purchasing products
Proposal to convert
invoice amount to
EMI’s
Customer Details
fed into the system
The algorithm developed will return fraud score based on inputs
The algorithm developed
will return fraud score
based on inputs
Instant mode
Approvals are made
instantly within 30 min
Normal Mode
Approvals are after
rigorous verification
Medium Risk
Feedback Process
FeedbackLoop
18. ROI of Modeling Exercise
www.valiancesolutions.com
Substantial decrease in loan
disbursement to fraudulent cases at
Point of Sale
Almost 6% of the originations are
referred to ‘Normal process’ in
which the fraud incidence is as high
as 5% which translates into a gross
saving of almost 1.5 million USD i.e.
50% of the VaR
Substantial decrease in the third
party cost of loan amount recovery
from the fraudulent cases.
Fraud Model led to
substantial decrease in
fraud thus improving the
Bottom Line
19. Case Study : Monthly Sales Forecasting for Direct Seller
www.valiancesolutions.com
Company Profile
Project Profile
Technology Used
Business Need Solution Benefits to Customer
Location
United States
Industry
Retail
Project type
Sales Forecasting
R
ARIMA
Linear/Non-Linear
Regression
Neural Networks
Develop Sales Forecasting
Model for Monthly Sales
• To build Monthly forecasting
Model with high degree of
accuracy.
• Forecast Monthly sales for next
9-12 Months
Neural Network based monthly
sales forecasting algorithm.
• Sales in last 1 year plus
external factors as inputs.
Model
Techniques
Error
Moving Average
And Exp
Smoothening
47%
ARIMA 32%
Linear
Regression
20%
**Neural
Networks
6%
20. Case Study Details
www.valiancesolutions.com
To identify Seasonal patterns and
factor affecting monthly sales.
Segment Agent workforce, to
improve forecasting Accuracy.
Forecast Monthly Sales for next 9-
12 months.
Quantitative Analysis of Monthly Sales Trend
Objective: Develop Sales Forecasting Model for Monthly Sales
To build Monthly forecasting Model with high degree
of accuracy.
Forecast Monthly sales for next 9-12 Months
21. Forecasting Solution
www.valiancesolutions.com
Monthly Sales
Raw Data
Sales Lag
Creation for last
12 Months
Train Neural
Network
Forecast for next
6 Months &
Calculate Error
Optimize Network
Weights
Forecast Sales
for Next 12
Months
• Various Forecasting Techniques are
used and best results are selected.
• Neural Network use Single Hidden
Layer Network with 24 Neurons.
Feedback Process