Picture a relationship. Six months in, the calls get shorter. Then the replies slow down. Then they start cancelling plans. You see all of it happening. But when it ends, somehow you still say, "I never saw this coming."
Businesses do this every single quarter. A customer's behaviour shifts. Their engagement drops. Their spend changes. The signals are loud. But nobody is watching. And when that customer walks out, cancels the subscription, switches the provider, churns. The boardroom acts surprised.
Today, using the IBM Telco Customer Churn dataset with 7,043 customers in IBM's public sample dataset, I want to show you exactly what those signals look like. Because once you see them, you cannot unsee them.
"A business that loses a customer without warning didn't lack data. It lacked attention."
First: how bad is the problem?
Of 7,043 telecom customers in this dataset, 1,869 churned ,that's a 26.5% churn rate. For every four customers this company has, it's losing one. And those 1,869 customers were paying an average of $74 a month. That's over $139,000 in monthly recurring revenue gone in this dataset's observation period.
That's not a customer service problem. That's a pattern recognition problem.
The contract is the commitment.
In a relationship, month-to-month means no commitment. No strings attached. No reason to stay if something better comes along. The same is true in business.
Customers on month-to-month contracts churn at 42.7%. That's nearly one in two. Customers on a two-year contract? 2.8%. Customers choose their own contracts, so some of this gap reflects who picks a two-year plan. Even so, contract type is a strong early warning sign.
The first 90 days are everything.
Think about how relationships end. Most of them don't collapse after years, they fizzle in the first few months, before real investment has been made. You never quite settled in. You never fully committed. And so when friction arrives, leaving is easy.
Customers in their first three months churn at 56.8%. Over half. By the time a customer has been with you for five years, that rate is 6.6%. The pattern is clear: every month a customer stays, they become less likely to leave. But you have to survive the beginning.
Tenure bands count customers with at least one month of tenure, and 'five years or more' means over 60 months.
"Over half of customers in their first 3 months churned. That's not just churn. That's a broken onboarding experience."
How you pay says how you feel.
There's a version of this in relationships too. When someone stops showing up automatically, stops making plans without being asked, you notice. The effort disappears.
Customers paying by electronic check churn at 45.3%. But customers on automatic bank transfer or automatic credit card? 16.7% and 15.2% respectively. The payment method isn't just a billing preference. It's a signal of how embedded a customer is in your ecosystem.
Mailed check, also a manual method, churns at 19.1%, so the effect is specific to electronic checks.
Manual payment = low friction to cancel. Automatic payment = the customer has made a decision to stay on autopilot. That decision matters.
The more invested they are, the harder it is to leave.
A customer using one service can cancel in 30 seconds. A customer using seven services has to find alternatives for all seven before they can leave. Depth of relationship creates switching costs and switching costs create loyalty.
Customers using just one service churn at 21.7%. Customers using all seven services? 5.8%. Churn is highest at two services (43.5%) and falls steadily from there to 5.8% at seven. Customers with one service are an exception: most of them are phone-only and churn less than customers with two.
When the signs align, the exit is almost certain.
None of these signals exist in isolation. The real insight in churn analysis isn't individual risk factors, it's what happens when they stack. In this dataset, there's a specific combination that should set off every alarm in your retention team:
Month-to-month contract + Fiber optic internet + No online security.
These are 1,774 customers. Their churn rate is 58.1%. More than half of them are going to leave. And the business can probably see every one of them in their CRM right now.
Can a model find them?
I trained a logistic regression on 80% of the customers and tested it on the other 20%, which the model never saw. Across five different splits of the data it scored about 0.84 on ROC-AUC, so the result is stable. When I rank the test customers by predicted risk, the top 20% contain 49.5% of everyone who churned. A random 20% would contain 20%. The strongest signals were tenure, contract type and internet service, the same ones this story has been following. The model predicts who is likely to leave. It does not show why they leave.
The signs were always there.
The breakup didn't come out of nowhere. It never does. Looking at this data, a business with a proper retention strategy should be watching for every single one of these signals — in real time:
- Customers in their first 90 days on month-to-month contracts — the highest-risk cohort in the dataset.
- Any customer matching the three-factor warning profile — prioritise for a personal retention touchpoint.
Customer churn is not a mystery. It is a measurement problem. The data tells you who is at risk, when they're at risk, and sometimes even why. The only question is whether you're paying attention — or whether you're going to look up in Q3, see the numbers, and say: "I never saw this coming."