Building a customer churn model

Ben:

Hi, this is Ben Rigby, CEO of Sparked, here with Cody Chapman, our Lead Data Scientist, and we’re going to talk today about building a customer churn model. Exciting stuff. Cody, we’ve written four phases up on the board here.

Phase one is defining churn. What are some of the possibilities?

 

Cody:

First, it’s really important that we have a clear definition of churn. If we’re a subscription business, this is very easy. We can just look at the contracts data to see if an account has renewed or churned at any given point. This is the simplest case.

A more complex definition is what we call “Activity Churn.” If you’re an Ecommerce business or mobile app, for example, you can label an account as “churned” if you haven’t seen it during a given period of time: if you haven’t seen it in six months, for example. Defining churn in this way is trickier, because the customer could come back at any point. But you just have to pick a timeframe that’s beyond when most customers have returned historically.

 

Ben:

Okay, so two ways to define churn: subscription-based and activity-based.

 

Cody:

Exactly.

 

Ben:

Phase two: what kind of data is interesting for us to look at when building a churn model?

 

Cody:

When we’re first building a model, it’s interesting to look at all of the data: billing data, usage data, and static account data such as what country they’re from, what type of plan they’re on, all of this different data. Then you’d use your domain expertise to prune it down into a more manageable set of signals. We don’t want to be feeding thousands and thousands of signals into the machine learning models. We can do that, but it’s more useful to prune in advance.

 

Ben:

Okay, so it’s really not a kitchen sink approach. It’s more a matter of feeding in data that we think will be potentially predictive?

 

Cody:

Exactly. It’s a combination of using your business expertise to define a potentially significant set of factors and then allowing the model to pick out what is actually statistically significant from that set.

 

Ben:

Okay, phase three: building a classification model. What are some of the possibilities?

 

Cody:

Yeah, so now we have churn definition and the pruned set of data. We then feed this data into a model that learns which factors are significant to churn. Black Box models tend to do a little bit of magic behind the scenes and give you a nice prediction out the back, but we don’t know exactly what’s happening in the process. Examples of these are support vector machines (SVMs) and gradient boosting machines (GBMs). There are a number of possibilities here.

 

 

An alternative is to use a more interpretable method such as logistic regression. It’s more transparent. It’ll give you a human-understandable result.

 

Ben:

Okay, so we got the data, we got definition of churn and we’ve chosen our classification model. What’s the final step?

 

Cody:

Now that we’ve chosen our classification model, we feed all of our data in and we let the model learn. And then we can use it for the fun part, which is the forecasting. The model will analyze the data from each account and will forecast an individualized churn probability for each account.

 

Ben:

Okay, so at the end of the day, we have a list of all of the accounts or all of the customers with a predicted risk score?

 

Cody:

That’s exactly right. How likely they are to churn at their next opportunity.

 

Ben:

Okay, and those are the basics for building a customer churn model. Thanks, Cody.

 

Cody:

Thank you.

 

Ben:

All right. See ya.

 

The following two tabs change content below.

Ben

Ben started his career as Lead Engineer at a social software company, acquired by Student Advantage, and co-founded a company that built award-winning web sites for Nokia, The North Face, Sony Pictures, and Calvin Klein. In 2002, Ben became CTO of DFILM, a web and mobile company with clients such as Sam Adams, Hyundai, Old Navy, IBM, The Sierra Club, and Scion. Ben graduated from Stanford University with Honors, Distinction, and Phi Beta Kappa.

Latest posts by Ben (see all)

0 replies

Leave a Reply

Want to join the discussion?
Feel free to contribute!

Leave a Reply

Your email address will not be published. Required fields are marked *