Churn Prediction Model
Category: Best use of Data Analytics
Churn Prediction Model Karl Vestli - EVP Consumer Sales, [email protected] Stig Breyholtz - Head of Analytics, [email protected]
How to know, in advance, and with great precision, which customers will churn and stay…? The story of how Aftenposten made it possible to predict churn by close to 90 % and reduce churn rates by 5 %
Our objectives By using relevant data and data analytics capabilities, we wanted to build a churn prediction model so that we could achieve the following: 1. Predict in advance, and with high precision, which customers are going to churn the coming period
2. Identify which customers to target with retention measures 3. Identify what specific actions that would have the greatest retention impact on each customer, or on defined customer segments
The churn prediction model is based on data from several different data sources
Customer data
+
CRM data
+
Digital usage data
+
3rd party data
Examples:
Examples:
Examples:
Examples:
• • • • • •
• • • •
• • • • •
• • • • •
Age Gender Geography Subscription type Subscription period Payment method
Contacts Channels Complaints Response data
Frequency Volume Time spent Products used Devices used
Income Wealth Education Housing Family situation
The model is built in R Studio and consists of a series of statistical analysis •
Descriptive analysis – in order to describe the effect on each parameter (or group of parameters) separately.
•
Survival analysis – to understand where in the lifecycle our customers tend to churn on a general basis.
•
Logistic regressions – to test the parameters, and their relative relevance and strength.
•
Factor and cluster analysis – to segment customers based on their respective churn drivers.
Conceptual overview and description of the model Data sources
Prediction model
Churn score
Clustering and segmentation
Actual Churn
1 Data gets pre-
2 Data runs through descriptive
3 Customers are scored
4 Churn segments are identified
processed, aggregated and loaded into the prediction model.
and predictive analysis.
(likeliness to churn) based on data analysis and predictions.
and further evaluated and clustered by churn drivers.
Actual churn data is monitored and fed back to evaluate and improve the model.
Based on the predicted churn score, all subscribers are classified into churn segments Likely to churn
Example (digital subscribers): Churn score Low Medium Low Medium High High
Segment Will not churn Will probably not churn Maybe churn Likely to churn
Maybe churn Share of Base 20 % 63 % 7% 10 %
Will probably not churn
Will not churn
Segments at risk are further analyzed and clustered by churn drivers By identifying the churn driver(s) relative to each customer or customer segment, the most effective retention activities can then be provided for each customer or customer segment.
Examples of retention activities in use - based on identified churn drivers Low digital engagement • Campaigns to sign up for digital accounts/login • Information about apps/products that are not used • Newsletters – most read articles last week
Manual payment combined with short renewal periods • Campaigns to encourage longer periods • Campaigns to change from manual to automatic payment methods (e.g. credit card) • New sales on longer periods and higher share of automated payment methods (reduce future churn) Days from purchase (renewal points) • Remind customers of value included in their subscriptions in time before renewal • Information of loyalty offerings • Update credit card (before expiry date)
Results?
Within the first year, the model predicts active and churned customers on a high precision level Example (full week print subscribers):
Predicted Actual Churn Active Churn Active
Active Active Churned Churned
Share of model base 6% 51 % 35 % 8%
• •
89 % of actual active subscribers were correctly predicted as active 81 % of actual churned subscribers were correctly predicted as churned
Overall predictability: 86
%
Monthly churn rate across all subscriptions reduced by 5 % from 2014 to 2015 Significant improvement from 2014 to 2015 throughout the year. More structured work with retention activities plays a role. Churn rates varies across subscription types due to different base and sales volumes. However, we are happy to see clear positive trends across all subscription types.
Next steps…
We will continue to improve and develop the churn prediction model in 2016 •
More data – we will provide the model with more behavioral data, e.g. richer data on digital engagement and response data from CRM.
•
Automated processing and scoring – we would like the churn predictions and churn scoring to be automatic and more frequent, and then spend more time on trimming and optimizing the model.
•
Real time usage/automated retention activities – we would like to use the churn prediction for real time decision making and automated retention activities when our subscribers interact with us in different channels.
Thank you! Category: Best use of Data Analytics
Churn Prediction Model Karl Vestli - EVP Consumer Sales, [email protected] Stig Breyholtz - Head of Analytics, [email protected]
Churn Prediction Model Karl Vestli - EVP Consumer Sales, [email protected] Stig Breyholtz - Head of Analytics, [email protected]
How to know, in advance, and with great precision, which customers will churn and stay…? The story of how Aftenposten made it possible to predict churn by close to 90 % and reduce churn rates by 5 %
Our objectives By using relevant data and data analytics capabilities, we wanted to build a churn prediction model so that we could achieve the following: 1. Predict in advance, and with high precision, which customers are going to churn the coming period
2. Identify which customers to target with retention measures 3. Identify what specific actions that would have the greatest retention impact on each customer, or on defined customer segments
The churn prediction model is based on data from several different data sources
Customer data
+
CRM data
+
Digital usage data
+
3rd party data
Examples:
Examples:
Examples:
Examples:
• • • • • •
• • • •
• • • • •
• • • • •
Age Gender Geography Subscription type Subscription period Payment method
Contacts Channels Complaints Response data
Frequency Volume Time spent Products used Devices used
Income Wealth Education Housing Family situation
The model is built in R Studio and consists of a series of statistical analysis •
Descriptive analysis – in order to describe the effect on each parameter (or group of parameters) separately.
•
Survival analysis – to understand where in the lifecycle our customers tend to churn on a general basis.
•
Logistic regressions – to test the parameters, and their relative relevance and strength.
•
Factor and cluster analysis – to segment customers based on their respective churn drivers.
Conceptual overview and description of the model Data sources
Prediction model
Churn score
Clustering and segmentation
Actual Churn
1 Data gets pre-
2 Data runs through descriptive
3 Customers are scored
4 Churn segments are identified
processed, aggregated and loaded into the prediction model.
and predictive analysis.
(likeliness to churn) based on data analysis and predictions.
and further evaluated and clustered by churn drivers.
Actual churn data is monitored and fed back to evaluate and improve the model.
Based on the predicted churn score, all subscribers are classified into churn segments Likely to churn
Example (digital subscribers): Churn score Low Medium Low Medium High High
Segment Will not churn Will probably not churn Maybe churn Likely to churn
Maybe churn Share of Base 20 % 63 % 7% 10 %
Will probably not churn
Will not churn
Segments at risk are further analyzed and clustered by churn drivers By identifying the churn driver(s) relative to each customer or customer segment, the most effective retention activities can then be provided for each customer or customer segment.
Examples of retention activities in use - based on identified churn drivers Low digital engagement • Campaigns to sign up for digital accounts/login • Information about apps/products that are not used • Newsletters – most read articles last week
Manual payment combined with short renewal periods • Campaigns to encourage longer periods • Campaigns to change from manual to automatic payment methods (e.g. credit card) • New sales on longer periods and higher share of automated payment methods (reduce future churn) Days from purchase (renewal points) • Remind customers of value included in their subscriptions in time before renewal • Information of loyalty offerings • Update credit card (before expiry date)
Results?
Within the first year, the model predicts active and churned customers on a high precision level Example (full week print subscribers):
Predicted Actual Churn Active Churn Active
Active Active Churned Churned
Share of model base 6% 51 % 35 % 8%
• •
89 % of actual active subscribers were correctly predicted as active 81 % of actual churned subscribers were correctly predicted as churned
Overall predictability: 86
%
Monthly churn rate across all subscriptions reduced by 5 % from 2014 to 2015 Significant improvement from 2014 to 2015 throughout the year. More structured work with retention activities plays a role. Churn rates varies across subscription types due to different base and sales volumes. However, we are happy to see clear positive trends across all subscription types.
Next steps…
We will continue to improve and develop the churn prediction model in 2016 •
More data – we will provide the model with more behavioral data, e.g. richer data on digital engagement and response data from CRM.
•
Automated processing and scoring – we would like the churn predictions and churn scoring to be automatic and more frequent, and then spend more time on trimming and optimizing the model.
•
Real time usage/automated retention activities – we would like to use the churn prediction for real time decision making and automated retention activities when our subscribers interact with us in different channels.
Thank you! Category: Best use of Data Analytics
Churn Prediction Model Karl Vestli - EVP Consumer Sales, [email protected] Stig Breyholtz - Head of Analytics, [email protected]
