## Predictive Churn Rate Modeling

by fabio on 15/08/2013**What is Churn Rate?**

Churn rate is a measure of **customer attrition**. It is basically defined as the number of customers who discontinue a service during a specified time period divided by the average total number of customers over that same time period. In the most cases, churn rate refers to product or services based upon a contract/fee business model with recurrent purchase, hence telecommunication, insurance, banking etc.etc. Churn rate is essential in order to calculate subsequent **KPIs live Customer lifetime value**, **ROI and marketing mix modeling**, every contract-based business has to be able to calculate this at any time.

**Why use modeling instead of actually tracking customer’s behavior?**

Churn rate modelling can be used in those instances in which there’s no technical way in order to actually track their behaviour or it is actually possible, but a time lag between the information being sent from the users and its collection and crunching can present some delay, so that the insight acquired wouldn’t be useful anymore.

The biggest benefit of relying on a churn rate modeling is that provides a predictive framework in order to take a look to the future and do some more predtiction work for other related KPIs, from total revenue to ROI and Customer Lifetime Value.

So far churn modeling has been focused on data crunching and correlation: providing

**Why predictive modeling works?**

Working predictive churn modeling work when we’re able to find the right one for the right business. There is no magic involved, there’s plenty of models out there and you have to find the right one, that one that fits your specific business case or your industry. Models based upon scientific phenomenas mainly works because those patterns are already being studied in alternative contexts, hence providing a good scientific footing for any upcoming theory.

** Uranium Decadence equation for churn modeling**

With this model we tried to figure out how to predict installation and uninstallation of a certain software we distribuited. The churn rate was meant to grow as installation grows but in a completely linear fashion.

The equation for the U238 decadence is a linear one:

*Uranium – 238 –> Pb – 206 + Alpha + Beta*

Starting off from a 6 months data set of installation and uninstallations, we run a test with the equation and the data that was available at that time:

The result was pretty impressive: the equation was returning a pretty realiable churn rate over the base line with a 5 to 8% tolerance margin, pretty accurate for being just a random model.

Now we’re able to base more predictive work on this number since we have a model that we proved that work, so all the subsequent KPI can be based on some sort of reliable ground, even inhereting the 8% tolerance margin.

There’s plenty of opportunities out there for models like this. The won’r solve your lack of tracking problem, still you can exploit them for the time being and, once proven to be effective, they can be used to have a look towards the future and run the same prediction on other relevant kpis.