Customer Analytics
In my role as a Campaign Analyst at an A.I. Ad Tech startup, I take present data to make predictions about future outcomes. The case here is an example of how I would predict churn rate for recurring customers.
* Since some data and information are confidential to the company and client, I can present only outlines and non-specific data representations.
Customer Churn Rate In Business Analytics
The ability to predict the who when how why of a customer's interactions is valuable for every business with returning customers. Churn rate is one of the key metrics, defined as the number of customers cancelling within a time period over the number of active customers at the start of that period. In order to apply a modeling technique to predict churn, we need to understand the customer behavior and characteristics which signal the risk of customer churn.
1- Characteristics
First step to churn analysis is identifying and selecting characteristics that could contribute to and reliably predict churn.
These characteristics could be categorized for ease of analysis, for instance:
-
Customer characteristics
-
Age
-
Location
-
Interests
-
etc.
-
-
Behavioral characteristics
-
Average amount of time spent on the website/app
-
Number of clicks/impressions before the final install
-
Time from install to sign up
-
etc.
-
2- Customer Lifetime Value
Another calculation which supports the churn prediction is to calculate the customer lifetime value (CLV or LTV). The CLV is a prediction of the total value generated by a customer in the past and in the future.
How To Predict Churn?
Having collected and selected some data that are correlated to customer churn rate, I developed a statistical model that calculates the churn probability for each customer. This model uses the selected characteristics and the CLV and expected lifetime. Each element is weighted such that the accuracy of churning is maximized. After developing the model, it is applied to all customers to calculate the likelihood of churn for each customer. Ranking the results then gives the top X customers who are most likely to churn.
You can typically have as many characteristics as possible (for example 50), but of those you will select out the ones with the highest prediction power.
