New RFM: Customer Retention in “Subscription” Businesses

Jim answers questions from fellow Drillers
(More questions with answers here, Work Overview here, Index of concepts here)

Topic Overview

Hi again folks, Jim Novo here.

How do you measure likelihood of customer defection when purchase behavior is highly orchestrated or executed due to repetitive billings? Yea, it’s a bit more complicated because “orders” really can’t express any kind of behavioral change, can they? So, you have to find indicators other than sales to provide the triggers. The Drillin’ the Drillin’ …

Q:  Jim, first let me say that I am enjoying your book VERY MUCH!!  Nicely done, and a nice job of integrating it with the CRM paradigm, 1-to-1 etc… I’m reading very slowly and finished the Latency Metric Toolkit.

A:  Great!  Thanks for the kind words.

Q:  I had a couple of questions on the Latency toolkit and the Latency tripwire, especially as it applies to environments with built in cycles for repeat purchases.

I am in a business where our resources are quarterly based, i.e. customers purchase our resource use them for a quarter and re-purchase the next quarter’s resource.  That is, we have a built in pattern, where customers would purchase our resources each quarter.  I was wondering how well I can use Latency with this type of built in cycle or if I would have any problems applying your Latency concepts to it, maybe they apply that much more readily?   In our case we try to call most folks who haven’t purchased within 2 weeks of a new quarter beginning.

A:  Right, a subscription-type business.  This is also an issue with utilities and other like businesses who bill about the same amount each month or have contracts for service (like wireless).  The answer is if the revenue generation really doesn’t represent anything to do with the behavior, then you simply look for other parameters to profile.  For example, a friend of mine was responsible for analyzing the likelihood of subscription renewal in a business that provided the content online.   Increasing Latency of visit was a warning flag for pending defection, and they triggered their most profitable campaigns based on last visit Recency.  In wireless, the correlations are found in payment Latency and age of phone.

Q:  Also, we have actually built a model with only 3 behavioral variables in it to predict those who would leave before they had actually left so we could do something about it.  The difference here from just calling folks would be that we are predicting who is at risk of leaving (defined by someone who hasn’t purchased 2 consecutive quarters).  

From the score, the model lets us sort our customers from hi to low and then we usually take the first 6 groups out of 20 to actually run our test groups (we do use our control groups!).  We usually run the model about 4 weeks prior to a quarter starting to flag the at risk customers, and then run our campaign. One of our questions is timing our use of this where we have a built in cycle of purchasing.

A:  Well geesh, what are you asking me questions for?  You already built it!  And where
models are concerned, simple is good!  You obviously learned some things from the book!

Q:  One of the variables was a Latency measure, but instead of looking at the overall customer average and comparing variations from that, we measured each customer’s history of purchasing and created a cycle-percent variable that in essence measured what percent they were over their expected time to buy and thus the model incorporated this “customized measure” for each person.  I was wondering your thoughts on this type of measure, compared to the overall customer average of buying cycles in your “Hair Salon” chapter, and also any thoughts on running an anti-defection campaign like this in an environment where we have a built in cycle for repeat purchasing?

A:  Sounds like a good model to me, a bit more advanced than the average person could swallow but you’re essentially using the customer’s own behavior to set the Latency tripwire for the customer.  Perfect for this kind of business, as long as you have enough behavior history on a customer to be predictive.  If you don’t (first subscription), you could always default to the average.

As far as campaign timing / anti-defection goes, if I understand the situation correctly, you probably want to time it back from the renewal event, e.g. test dropping the campaign 1 week before renewal, 2 weeks before renewal, 3 weeks before renewal, etc.  Once you get baseline for this, then matrix against “persistence” or your likelihood to renew model and try different offers, if you can.

So, for example, you will probably find out that people who are highly likely to renew require less lead time and perhaps no discount to encourage renewal; conversely, the less likely someone is to renew, the more lead time they need and the higher the discount.  And likely to renew probably roughly correlates with the number of past renewals.  So you end up with some kind of Frequency / Latency matrix that drives campaign timing and offer.

But don’t try to guess, test !!! Use of control groups will prove best path – even to CFO!

Get the book at

Find Out Specifically What is in the Book

Learn Customer Marketing Concepts and Metrics (site article list)

Download the first 9 chapters of the Drilling Down book: PDF 

Leave a Reply

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

This site uses Akismet to reduce spam. Learn how your comment data is processed.