Customer attrition is a bad thing. Businesses make significant investments in acquiring customers and losing them hurts!


But what exactly is attrition? Can we measure it? And if we can’t measure it, how can we possibly predict it?

Let’s start with a simple definition: a customer has attrited if she has stopped doing business with you.

For businesses that have contractual relationships with their customers, this definition is good enough. Your mobile provider has a contractual relationship with you. When the contract ends and you don’t renew, you have attrited. When you call and cancel, you have attrited. For these lucky businesses, it is simple to check if a customer has attrited and when it happened.

It is very different for retailers, however. Customers don’t have contractual relationships with The Gap or Best Buy. They can stop shopping with the retailer anytime they please. They may fall out of love with the retailer abruptly or gradually. Anything is possible.

In this setting, how can we even tell if a customer has attrited?

The extreme cases are easy. If you are an apparel retailer and a customer hasn’t shopped for 3 years, she has attrited (deal with it).

In general, however, we can’t tell with 100% certainty that she has attrited. We have to use the data at our disposal to infer what might be happening.

Let’s look at an example. This customer shopped a month ago. Is she attriting?


Well, we can’t say without more information, right? Here are her two previous purchases:


What about now? Well, while we can never be sure, we will likely guess that she hasn’t attrited. After all, it isn’t time for her next purchase yet so no need to worry.

But what if her prior purchases were:


We are in trouble! Something has changed for the worse in the relationship. From just eye-balling the time series, it looks like the customer has “missed” either two or three purchases. Again, while we can’t be 100% sure (hey, maybe she is backpacking in Europe and will resume shopping as soon as she is back), we will likely conclude that the customer has attrited.

Intuitively, this is what we are doing: if a customer “misses” her expected-next-purchase-time, we begin to get worried. The probability she has attrited starts to climb. As time passes and she doesn’t return, we get more certain that she has attrited.

In my artificially-simple examples above, it wasn’t hard to see what may be happening with the customer’s relationship. With real customer data, it is never this clean or easy. The average number of purchases per customer for retailers outside the grocery/drugstore vertical is typically in the low single digits. To glean a customer’s “natural” purchase frequency in this setting is very tricky. Further, there are many more variables we need to take into account because they may have valuable clues to help us figure out the health of the relationship and when the next purchase might happen e.g., the customer bought from a category she has never bought before (positive), the customer returned a number of items in contrast to her usual behavior (negative), lodged a complaint in the toll-free line about a recent store experience (very negative), posted a glowing product review (very positive).

Bottomline: you need to think about and model the problem in a rigorous, probabilistic manner using all the data you can get your hands on. If you do so, you can reliably predict if a customer is attriting and (very importantly) when the attrition began. You should refresh the attrition prediction daily and trigger an intervention (e.g., a win-back email offer) when a sudden increase in the attrition probability appears on your attrition radar.

The CQuotient platform does this out-of-the-box and we are powering reactivation campaigns using this approach.