If you sell products outside of a subscription, customer churn can be hard to identify. You can only guess that a customer is lost after you haven’t seen them in a while. Without a subscription, there is no good way to tell how much revenue you should expect from a given customer over the next year, or how many orders they might make. You can only guess at both, and as customer purchase behavior has evolved, so, too, has marketing strategy — guessing isn’t enough anymore.
At WhatCounts we use a mixture probability model to make these predictions. An active customer’s chance of making another purchase is governed by a two-parameter probability distribution. Their chance of becoming inactive — that is, churning and not buying from you anymore — is also governed by a two-parameter probability distribution. Intuitively, this mixture of probability distributions works as follows:
If a customer was acquired a while ago they’ve only made a few rare purchases since then, and you haven’t seen them in a while, their probability of having churned is low, as is their predicted revenue due to their low number of predicted future orders if they are, in fact, still active.
On the other hand, a customer acquired equally long ago who used to be a frequent shopper but has not been seen for a while has a high estimated churn probability. However, their predicted revenue will not be very high, because this high estimated churn probability translates in a low expected number of transactions.
In other words, your future revenue takes a hit from still-active customers who do not shop often and from frequent shoppers who are no longer active.
We can estimate the above-mentioned set of four parameters with varying degrees of precision, depending on your customer base (larger is better), length of observed history (longer is better), and share of repeat customers in the total (greater is better). These estimates allow us to make predictions for individual customers, but this is not their best use. They should instead be used for targeting groups. You can expect them to do two things well:
Target the right groups at the right time. This will help you gauge the residual lifetime value of different groups, and inform the distribution of your marketing spend along your customers’ expected lifetimes. Outside of more informed Win Back campaigns that Predicted Churn Probability affords you, the data in the Predicted Customer Value Report lends itself to messaging your best customers based on their future potential instead of past behavior. It also exposes a large swath of customers that are not necessarily Win Back audiences but are still highly influenceable based on their churn probability.
Improve your segmentation efforts. Customers with estimated churn probabilities higher than 90% are likely gone, and those with estimated churn probabilities lower than 10% are likely still active. It’s generally true that if your marketing dollars are to do any good, there are two types of people you don’t want to spend them on — people who definitely won’t purchase again and people who definitely will.
You will instead aim them at the middle of the churn probability distribution. You can define this middle more narrowly or more generously to suit your goals and your risk tolerance.
Probability-based segmentation of churn risk applies predictive data to marketing strategy that was previously influenced only by intuition. Even if you did have the ability to tailor your campaigns based on a generic “not seen since” date and used your best judgment to determine whether or not a particular customer was a churn risk, the reliance on intuition makes for a suboptimal approach.
In today’s marketing environment, it’s not enough to know whether or not someone will churn. Probability-based segmentation of churn risk gives you a data-driven way to understand how much sooner a customer will churn relative to the rest of your customer base, enabling you to deliver the right message with the right product recommendation to the right customers at the right time in their lifecycle.