Keeping a customer is usually cheaper than finding a new one. The hard part is knowing who is about to leave while there is still time to do something. Churn prediction gives every customer a score for how likely they are to stop buying, so your team can focus on the right people.
First, decide what “stopped buying” means
For a subscription business this is simple, because a customer cancels. For a shop or a service business, nobody cancels. They just stop coming back. So you need a rule, such as “no purchase in 90 days” for a store where people usually buy every month.
Look at how often your regular customers buy and set the rule from that. Everything else depends on this definition, so agree on it before anything is built.
Start with simple rules
You do not always need machine learning. Many businesses get good results from three simple measures:
- Recency: how long since the customer last bought
- Frequency: how often they buy
- Value: how much they spend
Customers who used to buy often and have suddenly gone quiet are your first list to look at. A spreadsheet can do this, and it is a good test of whether acting on the list makes a difference.
When a model helps
A machine learning model becomes worth it when you have more customers than your team can review by hand, and enough history for the model to learn from. As a rough guide, that means at least a year of data and a few thousand customers, including some who have already left.
A model looks at many signals at once, which simple rules cannot. It might notice that customers who contact support twice in a month and stop opening your emails are far more likely to leave than either signal alone suggests.
The data you need
- Order history: dates, amounts and products for each customer
- Customer details: when they joined, where they are, which plan or segment they are in
- Support history: number of tickets or complaints
- Engagement: email opens, logins or app usage, if you track them
Most of this already sits in your shop platform, CRM or accounting software. The work is joining it together so each customer has one row of history.
How the model works, in plain terms
The model studies customers from the past and learns what the ones who left had in common before they left. It then checks your current customers for the same patterns and gives each one a score. A score of 90% means this customer looks a lot like the ones who left before.
- Your dataOrdersSupportEmail
- One profile eachFull history per customer
- Model learnsFrom customers who left
- Risk scoreFor every customer
One common mistake is feeding the model information it would not have in real life, like a field that is only filled in after a customer has already gone. The model then looks very accurate in testing and fails in practice. A careful build checks for this.
What to do with the scores
A score only helps if someone acts on it. Simple actions that work well:
- A personal email or call to your most valuable customers at high risk
- A win back offer for customers who have gone quiet
- Asking for feedback to find out what went wrong
To see if it is working, contact some high risk customers and leave a similar group alone. After a month or two, compare how many from each group kept buying.
The same idea works for employees too. Our HR attrition predictor scores staff by how likely they are to leave, so HR can talk to them early. Read more about our machine learning work, or book a free call and we will look at whether your data is ready.