Churn prediction

What churn prediction delivers

Churn prediction estimates which customers will leave before they do. The difference from the churn rate lies in timing: the rate counts who is gone, the forecast shows who is at risk. Only that allows doing something before the loss instead of accounting for it afterwards.

In practice the forecast is a probability per customer. It is not a statement about individuals but about patterns. Anyone confusing the two treats a customer as lost who never was.

Which attributes are telling

• Recency: how long since the last order. The strongest single attribute.
• Change of rhythm: someone who ordered every six weeks and has not for four months stands out more than someone who orders rarely anyway.
• Falling basket: declining amounts at the same frequency point to fading interest.
• Interaction: opening messages and visiting the shop, where that data exists.
• Complaints and returns: an unresolved problem is a strong signal, independent of every other value.

How the benefit is measured

The forecast alone has no value. It becomes useful only when an action follows, and the benefit of that action is the actual figure. Three values are telling.

First the number of customers contacted and the share who buy again afterwards. Second the comparison with a group that was not contacted, because otherwise it cannot be separated whether the contact worked or the customer would have returned anyway. Third the contribution margin of the recovered customers less the cost of contact. The last value decides whether the method carries, and it is often smaller than expected.

Why risk is not the same as willingness

The most important misconception sits at the core of the method: a high churn probability says nothing about whether the customer reacts to contact. Some customers left for a reason that cannot be changed any more, for instance because their need shifted. Others are unremarkable and would react to a friendly reminder.

It follows that a good forecast is not enough. The additional question is which customers are open to contact. In practice that is often overlooked, so effort flows into customers who are no longer reachable. Selecting by both values is more sensible: high risk and expected response.

A second point is timing. Contacted too early the reminder feels pushy, too late it arrives after the decision. The fitting moment lies between two expected orders, not after them. A third point is the cause: a customer staying away because of a delivery delay needs different contact than one whose need has changed.

Frequent questions

• Do I need AI for it? A simple attribute such as recency delivers usable results in the first step.
• From how many customers does it pay? From some thousands of customers, below that the statement is too uncertain.
• What is the most common mistake? Contacting without a comparison group, then the benefit cannot be shown.