6 and 12 Month Predictive Churn Models

BenB

I'm working on building predictive models for to better forecast 6 and 12 month churn.  I'm happy to keep everyone posted on progress and what triggers/hitorical data is used.  but also happy to hear what has worked for others. 

Willow Moellering

@Ben Bunting, what are you using as metrics?  This might be a great place to brainstorm ideas!

BenB

Brainstorm Time - so please ignore grammer and fomat.  just typing outloud here.

1st I'm collecting data on churned customers (not necessarily why they churned, but what we know about them.  I'm hoping to use information below I can learn who is more likely to churn and when regardless of things like a healthscore, the csm, etc....  

• Tier (based on ARR) 
• Client Maturity we break our clients into 4 groups of what stage they are in building SEO stratedgy (from just beginning to center of excellence). 
• Length of Partnership, 
• Subscription Level, 
• Industry, 
• Team Size, 
• Region, 
•  
• Add more please  

Next I will use data we have on existing clients to start creating triggers for churn and try to assign value: 

• Depth, Breath of adoption - are they using the service, and how much of the platform 
• Login Activity
• Internal Champion ? 
• Change in Primray Contact/Decision Maker/Executive
• Health Score (Our existing score is composed of a few of these already) 
• Bugs/Support Tickets
• Onboarding/implementatoin (completed on time? 
• Financial Help (paying and paying on time or on their typical schedule
•  
• Add more please

I dont think ill build a perfect model, but going through the process will certainly get my closer.  1 goal is to understand WHO will churcn, but beyond that i want to understand the amount of revenue i predict will churnc based on the model. 

Areya Dargahi

@Ben Bunting one thing I do encourage you to do (if not Initially, in the future), is look at your health scores situationally  or based on the lifecycle, e.g. on boarding score, go live score, etc

Just the most simple example here but if you use usage as a factor, since they’re not actually using the product during implementation, it could skew the score quite drastically  

Love this topic and I have to say, you’ve most certainly thought of some fantastic factors. 

 

Jeffrey Nadeau

I have a formula I use for predicting future churn that gives me a Best Case, Medium Case and Worst Case scenario. I use Low, Medium and High risk for each customer which then in the formula calculates the likelihood of renewal. I will say that predicting churn 6-12 months out is hard to do because so much can change. The calculator I've used does allow you to have a good range to work from to forecast future churn. 

Low Risk assumes 90% likelihood of ARR renewal

Medium Risk assumes 60-90% likelihood of ARR renewal

High Risk assumes 0-60% likelihood of ARR renewal

Vijaya Vardhan P

@Ben Bunting This is interesting (read: my fav stuff). What kind of model are you trying to build? I dabble in Python. I built a simple text analytics ML algo last year to put a number to customer sentiment from unstructured feedback. Nothing fancy. I used the existing Python API's. 

Thinking out loud: Can we develop a 'context' for churned customers to be able to define the 'why' behind the churn? I'll be watching this space.

BenB

Thanks Jeff - great suggestion!  And you're so right, change is happening all the time!  Thats why I was thinking id have better luck predicting churn of total ARR instead of predicting exactly which customer would churcn.  i could bucket total ARR of each Low, Medium, High and then apply expected churcn from each bucket.  I'll have to play around with it.

BenB

Agreed.  Our healthscores still need to be reviewed to get closer to a stronger #.   Some of our best customer live and die by our platform, BUT they use it very specifically.  So they may only have a few users, use a couple features.  Meaning some of their usages/adoption could be too low.   Its a moving target ...

Areya Dargahi

@Ben Bunting I don't know enough about your business but if you're able to segment those customers who use it "very specifically," you may consider a specific score for just them if your bandwidth allows for it. 

Jeffrey Nadeau

Makes sense and agreed. Definitely hard to predict specific customers who will churn but easier for ARR. Good luck as you build the model out!

Andy Barton

Interesting conversation... f you can build a churn predictive model, how does it help? Can you flip it slightly and create a model that identifies high risk 12 months out (depending on the size of your business and subscriptions, 6 months might be too late even)... 

Something I always use to help me guage risk, or have as part of health score is  to consider a customers willingness or participation in advocacy activities (reference calls/visits, case studies, success stories etc)... 

BenB

Thanks Andy - nice addition to potential signs or renewal/churn!   The predictive model certainly would help me better highlight high risk.  but overall it helps bc this is a key metric for me and details that i run up to ownership. Much like a salesVP needs to understand their forecast as far out as possible. 

Andy Barton

yes indeed @Ben Bunting . I agree when you are looking at predictability for overall ARR its good to have a model ! when you have the model created, will you share with the CSMs? If not, you could track their bottoms up forecast accuracy