Should They Stay or Should They Go?
The Prediction of Customer Churn in
Energy Sector
Michela Vezzoli (University of Milano-Bicocca)
&
Cristina Zogmaister (University of Milano-Bicocca)
Big Data in Psychology 2018
Churn Behaviour
Customer Lifetime
Acquisition Phase Retention Phase
Cross-Sell & Up-Sell
Profits
Acquisition
Time
Big Data in Psychology 2018
Churn Behaviour
Customer Lifetime
Acquisition Phase Retention Phase Churn Phase
Cross-Sell & Up-Sell
Profits
Acquisition Churn Win Back
Time
Big Data in Psychology 2018
Why study churn behaviour?
Economic reasons
Attracting new customers costs 5 to 6 times more than
retaining of the existing ones
Long-term customers generate more profits
Long-term customers are less sensitive to competitors’
marketing campaigns
Long-term customers are less costly to maintain over time
Long-term customers provide new referrals through positive
word-of-mouth
Big Data in Psychology 2018
Why study churn behaviour?
Psychological reasons
More importantly, churners are
unhappy, unsatisfied and no-
more-loyal customers
Big Data in Psychology 2018
How to study Churn: Making predictions
Stay
Historical
Information on Predictive Modelling Prediction
Customers
Go
Machine Learning
Big Data in Psychology 2018
The aims of the study
Shed light on
1 2 3
Develop the Understand
the
predictive predictive
consumer
model relationships
psychology
Big Data in Psychology 2018
Methodology for developing predictive
churn models
Predictive modelling turns data into information and
information into insight
It does not demand a priori hypotheses ➜ Data Driven
The Data Mining Process
Model
Data Data Model
Training and Action
Acquisition Cleaning Testing
Tuning
Big Data in Psychology 2018
Methodology for developing predictive
churn models
Predictive modelling turns data into information and
information into insight
It does not demand a priori hypotheses ➜ Data Driven
The Data Mining Process
Model
Data Data Model
Training and Action
Acquisition Cleaning
Tuning
Testing
Big Data in Psychology 2018
Methodology for developing predictive
churn models
Predictive modelling turns data into information and
information into insight
It does not demand a priori hypotheses ➜ Data Driven
The Data Mining Process
Data Model
Data Model
Training and Action
Acquisition Cleaning Tuning
Testing
Big Data in Psychology 2018
Methodology for developing predictive
churn models
Predictive modelling turns data into information and
information into insight
It does not demand a priori hypotheses ➜ Data Driven
The Data Mining Process
Model
Data Data Model
Training and Action
Acquisition Cleaning Testing
Tuning
Big Data in Psychology 2018
Methodology for developing predictive
churn models
Predictive modelling turns data into information and
information into insight
It does not demand a priori hypotheses ➜ Data Driven
The Data Mining Process
Model Model
Data Data
Training and Action
Acquisition Cleaning
Tuning Testing
Big Data in Psychology 2018
Methodology for developing predictive
churn models
Predictive modelling turns data into information and
information into insight
It does not demand a priori hypotheses ➜ Data Driven
The Data Mining Process
Model
Data Data Model
Acquisition Cleaning
Training and
Testing Action
Tuning
Big Data in Psychology 2018
Data acquisition and cleaning
Outliers Treatment
Handling Missing Values
820.123 81.813
Customers Defining Time Window Customers
Feature Selection
CRM Getting Know the Data Electricity
Residential
BILLING
Single POD
CAMEO
Original Datasets Data Cleaning Targeted Population
Predictors
Socio-Demographics
Age, Sex, Regional area
Account
Customer Type, Length of the contract, Acquisition Channel, Loyalty Program
Member, Payment method, Online Billing
Behavioural
Number of complaints, Number of change offer, Number of contacts, Number
of retention proposal, Contract starts with a transfer, Number of cross sell
proposal, Digital customer, Number of previously churn
Socio-Economics
Socio-economic status, Presence of adults over 60, Presence of children,
Household size, Education, Building age
Train – Test split
Training
Train Set Model
Data
Test Set Test model Performance
Total Dataset Training Set Test Set
N of Churner (%) 6899 (8.4%) 4848 (8.5%) 2050 (8.3%)
N of Non-Churner (%) 74915 (91.6%) 54421 (91.5%) 22494 (91.7%)
Total 81836 57269 24544
Class imbalance
Number of non-churners is far higher
than the number of churners
Resampling Approach: SMOTE
SMOTE Training Sample
N of Churner (%) 24240 (45.5 %)
N of Non-Churner (%) 29088 (54.5 %)
Total 53328
Modelling phase: Training and testing
Churn prediction is a supervised learning Decision Tree
Interpretability
classification task Logistic Regression
Support Vector Machine
Random Forest
Neural Network
Accuracy
Big Data in Psychology 2018
Modelling phase: Training and testing
Churn prediction is a supervised learning Am I hungry?
Yes No
classification task
Have I 25 €? Go to sleep
Yes No
Decision Tree (CART and C5.0)
Go to Buy
restaurant hamburger
Big Data in Psychology 2018
Modelling phase: Training and testing
Churn prediction is a supervised learning
classification task
Decision Tree (CART and C5.0)
Logistic Regression
Big Data in Psychology 2018
Modelling phase: Training and testing
Churn prediction is a supervised learning
1
classification task
True Positive Rate
0.5
Decision Tree (CART and C5.0)
Logistic Regression
0 0.5 1
False Positive Rate
Performance measure: Area Under the ROC Curve (AUC)
Range values from 0.5 (random model) to 1 (perfect model)
Big Data in Psychology 2018
Results
Model AUC Train AUC Test
CART Unbalance 0,51 0,51
CART + SMOTE 0,76 0,56
C5.0 Unbalance 0,51 0,51
C5.0 + SMOTE 0,73 0,56
Logistic Regression Unbalance 0,67 0,68
Logistic Regression + SMOTE 0,77 0,63
Big Data in Psychology 2018
Odds ratio: Socio-demographic
predictors
Center 1.17
North East 1.01
Regional Area
South 1.40
Islands 1.14
Sex Female 0.99
No relationship
Age 0.99 Decrease the likelihood of churn
Increase the likelihood of churn
Odds ratio: Account predictors
Agency 2.17
Counter 0.47
Acquisition
Call Center 1.09
Channel
Web 1.18
Tele-selling 1.25
Customer Type Dual 0.89 No relationship
Decrease the likelihood of churn
Payment Method RID 0.95 Increase the likelihood of churn
Odds ratio: Account predictors
Lenght of the Contract 0.986
On Line Billing Yes 0.84
Start with
Transfer
Yes 0.89
Loyalty Program
Member
Yes 0.78
No relationship
Digital Customer Yes 0.95
Decrease the likelihood of churn
Increase the likelihood of churn
Odds ratio: Behavioural predictors
Number of
Cross-Sell Proposal 0.70 Contacts
1.04
Number of
Change Offer 0.449 Complaints 1.36
Retention Proposal 1.478 Previously Churn 1.50
No relationship
Decrease the likelihood of churn
Increase the likelihood of churn
Odds ratio: Socio-economic predictors
Socio-Economic
1.08 Household Size 0.99
Status
Presence of adults
1.02 Building Age 0.97
over 60
Presence of children 1.02 Education 0.99
No relationship
Decrease the likelihood of churn
Increase the likelihood of churn
Discussion
Contributions to the knowledge on consumers’ churn behaviour
Implement machine learning techniques and data mining
methodology into the consumer psychology research
Logistic regression outperformed CART and C5.0 decision trees.
Moreover, the logistic regression has shown to be robust
Big Data in Psychology 2018
Limitations and Developments
Consider other predictors
Examine multiple retailers
Actual behaviours of electricity consumers ➜ Ecological Validity
Disentangle the causality of some of the effects we found
Big Data in Psychology 2018
