How to Build a Prediction-Based Customer Churn Program
Unlike most market research practices, using predictive analytics to address customer churn is a highly iterative process. Specifically, there are two iterative phases: building and refining your data set and model; and testing and learning into your response program. But before digging deeper into these processes, let’s first look at why anyone would even want to adopt a predictive approach to customer churn. A predictive customer retention program is one of the best ways you can drive tangible business results from your customer experience management program.
Why Predictive Modeling?
When it comes to using data to understand customer churn, there are two general approaches: driver analysis and predictive modeling. Driver analysis can be as simple as correlation analysis and as sophisticated as multiple regression - but either way, the output is the same: the identification and relative strength of leading indicators (independent variables) of customer churn. This type of output is referred to as top-down: it tells you in general and in aggregate what things contribute to customer flight. Top-down data is great for high-level strategic planning because an organization can easily prioritize what to focus on and where to allocate resources.
But unfortunately, we do not live in a one-size-fits-all world where people all act the same. And in our highly fragmented, personalized and digitized society, a top-down approach to understanding customer behavior does not provide individual customer-specific data. For this, we need predictive modeling. A predictive modeling approach to customer churn provides scores of likelihood to leave at an individual customer level. Because values are attached to each customer, predictive modeling can be referred to as bottom-up. With bottom-up data, you can specifically identify the risk of churn for every single customer. With data like this, you can then target your customer responses at an individual level.
How Can I Do Predictive Modeling?
It used to be that building powerful predictive models was only possible for people who were highly educated in statistical algorithms and comfortable in R or Python. But now, companies like Qualtrics (with their Predict iQ product) have made it possible for people without any statistical or programming background to build churn models quickly and inexpensively. Predictive modeling platforms are designed to do all the statistical heavy lifting, so you can do all the human strategic thinking. For example, Qualtrics’ Predict iQ platform is built on a learning system that identifies and builds on the complex interrelationships that exist between variables and their connection to churn. It deploys complex machine learning via neural network which identifies the relationships from data in a way that mimics how the human brain learns.
Where to Start: Collect Your Data
To generate your initial predictive model, you’ll need actual historical customer data. Like all insights endeavors, predictive modeling follows the “garbage in garbage out” rule. Unless you have good data that includes strong indicators of prediction for key customer segments, running your data through predictive modeling will not yield meaningful or actionable results. Here are the three types of data you need to consider to build your predictive model:
- Operational Data (O-data): This is the customer transactional data that probably sits with your BI or finance team (things like CustID, order size, tenure, average order value, products ordered, etc.)
- Experience Data (X-data): This is the customer feedback data that you collect when you directly survey your customers (e.g. NPS, satisfaction, likelihood for next purchase, etc.)
- Churn Data: This is a binary value (1/0 or yes/no) identifying whether or not the customer has churned. (Note: Identifying an individual customer’s churn is an easy task if you have a subscription or renewal business - because then you have a clear indicator or churn. But if you don’t, put some thought around how your business identifies customers who clearly will not come back.)
Once you have collected this data, run it through your predictive modeling platform and among other things it will assign a “likelihood to churn” score to each customer. At this point, you’re probably going to want to refine your model with more and different data in order to generate as accurate of a model as possible. When you have your model built and you can then feed new customer data into it in order to determine the churn scores for new customers as they come in.
So Now What? Start Testing Preventative Measures
Once you have a system in place that is assigning churn scores to your customers, you then enter phase 2 of the process: testing and learning into your response program. Since predictive scores are assigned to your customers at an individual level, you need to outline a methodical approach to testing into the most effective response system. Specifically, you will need to test:
- What to say: What messages or offers do you think can rescue churn-likely customers?
- When to say it (and how frequently): Soon after purchase, just before churning or at intervals along the way?
- Who does it impact the most: Which churn score ranges are most susceptible to your messages/offers?
The beauty of this type of testing is that you have a very clear indicator of success: an improved rate of churn. (Note: One easy way to inform “what to say” is to look at the results from your churn driver analysis for direction.)
Conclusion: Test, Learn, Refine & Repeat
As indicated, building a predictive churn program is an iterative process - it takes a lot of patience, creativity, and strategic thinking. And along the way you will garner some incredibly rich and actionable insights. But in the end, you will have a response system that will in fact measurably keep more of your customers coming back.
For a more in-depth guide to running a churn predictive modeling process, watch our on-demand webinar about predicting customer churn.
March 24, 2020
Asking your customers for reviews: 5 tips for receiving important user feedback
March 13, 2020
For the 234th time… why contact center silos mean frustrated customers
February 27, 2020