Customer Churn

Analysis and
Prediction
William Chen
Margaret Ji
Catherine Ruan

1
Overview

§ Use the consumer dataset to:

Goals § Segment the Fulton Bank customer base
§ Build a model that predicts customer churn

Agenda

1 Data processing

2 Customer segmentation

3 Feature scoring and predictive model implementation

4 Business recommendations
2
Data Processing

Objective § Prepare data for analysis by removing and modifying data

Numerical/Binary Categorical Balances Missing Totals

§ Keep columns § Find appropriate • Set missing § Fill blank cells with
containing level of detail balances to 0’s or -1’s
relevant § One-hot encode -10,000 depending on
characteristics of • Use smooth context
customer symmetric log
segments scaling

125 Columns
3
Customer Segmentation

Objective § Figure out if consumers naturally fall into certain groups

Methods Results

§ Dimensionality reduction
§ Finding the optimal number of segments Churn
§ Clustering
§ Segment analysis
Behaviors

4
Customer Segmentation

Objective § Find a more concise representation of data

Reconstructed Data vs. Actual

Method Percent of Data Described
Data
§ Dimensionality reduction
§ Autoencoder
§ Principal Component
Analysis
§ Performance analysis
§ Reconstruction loss

5
Customer Segmentation

Objective § Use unsupervised learning to segment customers into groups

Why unsupervised segmentation?

Select best number of clusters based on eigengaps
§ Cherry picking metrics may not capture
nuances in the data
§ Unsupervised clustering can cover as much
information as possible Perform large scale clustering using KMeans
§ To be understand churn, it is good to first
understand its correlation with consumer
behavior
§ Spectral clustering is best suited for nonconvex Examine clustering performance using elbow
geometry method

6
Customer Segmentation

Objective § Use unsupervised learning to segment customers into groups

Largest increases in eigenvalues Minimized inner cluster distance

7
Customer Segmentation

Segment Characteristics

Highest churn | Lower LTV | Lowest average mobile logins | Use Venmo/PayPal the
“Swing by”
least | Highest % of closed accounts | Lowest number of remote deposits
Lowest churn | Highest number of calls to call center | Highest average age in
“Loyal”
households | Highest number of saving accounts | Highest % of high income
Highest average mobile logins | Highest % Uber/Lyft payments | Use Venmo/PayPal
“Techy”
the most
Largest percentage of H-P H-F | Highest direct deposit amounts | Lowest amounts
“Valuable” of check deposits | Fulton customer the shortest | Younger households (often Gen
X) | Highest % of middle income
Highest average of billpay transactions | Highest average (deposit, investment, loan)
“Active”
products in household

8
Customer Segmentation

9
Customer Segmentation

1
0
Customer Segmentation

1
1
Business Recommendation: Target Customer Segments

§ Classify customers into segments to provide customers with targeted

Objective recommendations that meet their needs and increase loyalty

Acquisition Servicing Relationship Retention

§ Identify and execute § Provide services • Build in-depth • Prioritize converting

campaigns targeting tailored to the relationships with “swing by”
customers with customer’s needs customers via customers to other
characteristics of based on segment analytics-based segments with lower
low-churn segments traits personalized churn
§ Recommend or services
cross-sell products
associated with
loyalty

12
Feature Scoring Procedure

Objective § Which features are giving the most improvements to accuracy in a nonlinear model?

125 Choose 2 = 7750 Columns

Fit 7750 Random Forests

(Each column will be in 124 of the models)

Distribution of average accuracies for each model that a column participated in

Scored column as the equally weighted average of the mean, median, 90th percentile, and max of the
distribution

Graphed column scores and picked a natural cutoff point

13
Feature/Column Scoring Results

Top Scoring Columns Score by Sorted Column Index

Bottom Scoring Columns

14
Starting Product

Most vs. Least Likely to Churn Churn Fraction by Product

15
Predictive Model Implementation

Objective § Design a model that predicts whether a household churns or is kept

Method Implementation Notes

§ Nonlinear model allows for complex interaction

§ By using “class_weight = ‘balanced’” in the model, we make sure the accuracy on the kept households
and churned households are prioritized equally
§ Since there are fewer churned HH, the precision suffers, but this is in line with business intuition of losing
a customer is more expensive than the cost to keep an existing customer from churning

16
Churn Model Metrics

Accuracy-Kept Accuracy- Precision Recall F-score

HH Churned HH

Random Forest (sqrt) 0.867 0.94 0.656 0.94 0.773

Logistic Regression w/ L2 0.408 0.848 0.28 0.848 0.421
Random Forest (all) 0.792 0.991 0.564 0.991 0.719

F.S. Random Forest (sqrt) 0.866 0.933 0.655 0.933 0.769

F.S. Random Forest (all) 0.793 0.991 0.565 0.991 0.72
AdaBoost 0.988 0.92 0.954 0.92 0.936
F.S. Adaboost 0.983 0.922 0.935 0.922 0.928
§ Random forest “sqrt” vs “all” refers to checking sqrt(features) or all features at each split
§ F.S means the model is run on feature selected data - the top 35 rows
§ Logistic regression has poor performance, but we may be able to get more meaningful significance data
out of it. 17
Business Recommendation: Utilize Predictive Variables

§ Use variables most predictive of churn to inform insights and strategies personalized
Objective for the customer

Understand Predict Strategize

• Examine intuition for starting

• Ensure data is robust enough • Reconsider profitability of
product and other high-
to draw conclusions high-churn products
scoring variables
• Use a nonlinear model to • Encourage customers to
• Improve data tracking to
predict whether customers switch to or add low-churn
include more of customers
are likely to churn products
product profile

18
Conclusion

§ Use the consumer dataset

Goals § Segment the Fulton Bank customer base
§ Build a model that predicts customer churn

Recommendations

§ Customer segmentation
§ Use customer characteristics to segment customers, allowing for easier acquisition, servicing,
relationship development, and retention of customers
§ Predictive model
§ Predict the likelihood of churn in an individual customer
§ Formulate strategies based on a trait’s association with high or low churn

19