06-01-2017 11:22 PM
I am building the churn prediction model.
Most of my data consists of transactions (several rows for each customer) with dates: purchases, logins,
calls, etc. I also have a table that has a churn date for each customer.
Is there a way / algorithm that will make a prediction without me having to "flatten" the table to one row per customer?
I am asking not because of the amount of work, but because when flattening the data to measures I might miss out some important explanatory variable that I will not think about.
For example: let's say I will create a measure "Number of purchases in the last month" but it will not predict the churn. While the real sign of the forthcoming churn is the "decrease percentage of number of purchases in this month compared to two months before".... Or the real explaining variable would be the "weekly frequence of purchases", etc.
I hope you understand what I mean.
Is there a way to do it?
06-02-2017 12:00 AM
I've done data preparation for customer churn models in the past and you really do need to roll up your data to one row per customer. Please note that it's not only account behaviour that could drive churn, but also customer attributes as well, like age, education, employment, location, and so on.