As part of my final Capstone Project piece, I developed a customer retention strategy piece on how to improve customer churn and developed a recommendation strategy based on the analysis and insights from the data set.
3. Australian retail bank offers a range of financial
products and services to market
4th largest branch & ATM distribution networks
nationally
4 credit card tiers: Blue, Silver, Gold & Platinum
4. ???
Resulted in the
customer churn rate
of 16% in 2020
(increased from 4%
YoY)
Don’t the know
reasons behind why
customers leave
OurBank
1 2
5. 60% In the past six years alone, the cost of
acquiring new customers has increased
60% (Hubspot, 2021)
Acquisition marketing and new customer
efforts can cost anywhere between 5-25
times more than retaining existing
customers (Deloitte, 2019) 5-25x
6. 2020 saw a 16% increase
in customer churn, up 4%
from 2019. Establishing a
KPI metric of a 10% churn
rate over 2021 would mean
efforts are effective
01
CHURN RATE
REDUCTION
Being able to classify
customer behaviours on
how they engage and
interact with the OurBank
product/service
02
DIFFERENTIATING
FACTORS WHICH
DISTINGUISH
CHURNED AND
LOYAL AUDIENCES
03
IDENTIFY AREAS
WHERE ADDITIONAL
DATA POINTS CAN
BE COLLECTED
How will we measure program success?
Classify further data points
around the OurBank
customer lifecycle, other
related products,
touchpoints and qualitative
data
7. Improvements in collecting data
points to have a more integrated
view of the OurBank customer
Reduce customer churn with a 10%
churn target KPI for 2021
Pinpoint a machine learning model
to predict the likelihood to
remaining a loyal OurBank
customer or leaving the bank
1 2
Develop key strategies based on
data analysis, variables and their
relationship with customer
churn/existing
3 4
8. IDENTIFIED
PROBLEM &
OBJECTIVES
WRANGLE,
PREPARE
& CLEAN DATA
ANALYSE DATA
1
2
Exploratory Data Analysis (EDA)
Correlation Matrix
3
Clustering Predictive Analytics, K-means Clustering & Logistic Regression Modelling
4
Python language to streamline data set
Tableau & Excel Visualisations
9. 83.9%
16.1%
• Immediately, we see a limitation of imbalanced distribution of existing customers vs. churned
customers.
• This presents a challenge for predictive modelling as machine learning algorithms used for
classification assume an equal numbers of samples in each class
10. We found more females (53%) than men (47%) are using OurBank credit card services
Interestingly, with respect to both average balance available and average credit limit, the balance
for men are 3x greater than females.
53% 47%
11. Based on the sample size, we can see Blue credit cards make up 93% of our sample size of the
total share of existing customers, as well as churned customers. Again, is this a potential
organisational issue or an issue related to our sample?
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
Blue
Silver
Gold
Platinum
Credit Card Category Breakdown by customer type
Existing Customer Attrited Customer
93%
12. Total
Revolving
Balance
Summary of Analysis:
1) Imbalanced data set
2) There was no significant
differences between the segments
across demographic variables
• Gender
• Age
• Dependent count
• Education level
• Marital status
• Income category
3) Churned customers were less active
& less engaged with their product.
Significant variables of interest lie in
product/customer behavioural related
variables…. (Exhibit A in appendix)
Total
Transaction
Amount
Total
Transaction
Count
Transaction
Count Change
(Q4/Q1)
Card
Utilisation
Ratio
13. 1. Using credit card behaviours as a gage to launch new product offerings and
flagging risks for the bank
Intuitively, we can see as the
total revolving balance
increases (balance unpaid at
end of each cycle), average
utilisation ratio also increases
• Cluster 1: least risky for
the business to run tests,
trials, new product and
services offerings towards
• Cluster 2: stable and loyal
customers
• Cluster 3: revolvers!
14. 2. Benchmarking behavioural/transactional data for Loyalty and Retention Reward
Marketing Programs and Initiatives
Opportunities to review card tier structures, utilise
rewards and incentivise consumer and product
engagement through cashback tap-and-go
rewards to encourage engagement and activity
(similar to Westpac, HSBC and NAB in-market
offerings)
Using total transaction count or total
transaction amount as a benchmark for
understanding behaviours for Loyalty, Reward and
Retention programs and initiatives. E.g. knowing
total transaction counts >100
15. 3. Continuous customer experience improvements through efficient resource
allocation and servicing across all touchpoints
Churned customers made up to 6 contacts with OurBank
before churning, compared with 5 for existing customers –
WHY?
From an internal resourcing and achieving economies of scale
point of view, our are branches, contact centres and e-channel
workforces optimised and equipped to service different
tiers/products?
Given the nature of the already competitive Australian retail
banking industry, is it feasible to remain mass market or better
to capture a niche? (Brand realignment)
16. Collection of qualitative data to overlay quantitative data points. E.g.
NPS Customer Survey results vs. customer feedback /No. of contacts
within the past 12 months?
Develop triggers based on benchmarking metrics at the appropriate
time to communicate with customers based on transactional/behaviour
activity. E.g. reengagement campaign
Integrated 360 view of the customer – what other products/services do
they engage with, lifecycle stage, hobbies and interests?
Greater upsampling of the data set for machine learning accuracy.
Current model is a good measure based on 84-16% split, however will
skew results based on the data due to unequal category distributions
19. Exhibit B: Logistic Regression Modelling
• We had limitations to the supervised machine learning model and inaccuracies existing in the
model as there were unequal distribution of values between categories, i.e. churned and loyal
customers thus resulting in bias.
• In best managing the model moving forwards, recommendations are to upsample.
• The model is not feasible for use due to our current data set, without upsampling.