Churn Analysis

Within the financial services industry, where customers generally tend to stay with a company for a longer term, churning could lead to substantial revenue loss. 

You can identify customers who are likely to churn by making precise predictions, revealing customer segments and reasons for leaving, engaging with customers to improve communication and loyalty, calculating attrition rates, develop effective marketing campaigns to target customers and increase profitability. With advanced modeling algorithms and a wide array of state-of-the-art tools, you can develop powerful models that can aid in the accurate prediction of customer behavior and trends and avoid losing customers.

Voluntary Churn and Involuntary Churn

Finding variables that will help identify voluntary churn vs involuntary churn is important, but the data is difficult to collect. A customer may be: 

• cutting back on expenses because they are retiring
• graduating college and moving to another city
• searching for better prices

Service industries with a month-to-month subscription model (join the Gym and contact us to cancel) tend to have better luck collecting this data. Sometimes domain experts can help with a "rule of thumb" that identifies voluntary vs involuntary. Sometimes alternative variables can be used as signals of community health to help identify causes of voluntary vs involuntary churn. Customer zip codes can be linked to :

• census data 
• crime data
• economic data
• stock market
• weather data

When possible, data related to involuntary churn is removed before building a model. This allows the modeler to focus on issues that can be changed. 

Scoring Retention

Once the model is built, when and where will it be executed?

Will Janice in marketing execute the model against yesterday's data?

Will the model be executed when someone tries to cancel an order and then prompts the call center employee to ask a couple of questions? 

Will a nightly job score all existing customers for churn and notify the account manager to follow up? 

Will the model be executed in a Hadoop environment as health data streams in from devices that patients are wearing?