How to Predict Customer Churn From a Spreadsheet

Customer churn is when a customer stops buying from you — cancels a subscription, closes an account, or simply does not come back. Churn rate is the share of customers who leave in a given period:

churn rate = customers lost during the period ÷ customers at the start of the period

Churn prediction goes one step further: it uses your history of customers who stayed or left to estimate which current customers are most likely to leave next, so retention effort goes to them first.

Measured example: on 1,408 customers held back from training in the IBM Telco sample, the 10% the model ranked most at risk churned at 72%, against 25% across all customers — almost three times as likely. Making a retention offer to the top-ranked half was worth about $42,000 net, against $11,000 for making it to everyone, using the placeholder costs below.

What your spreadsheet needs

One row per customer, and one column saying whether that customer left (yes/no) during the period you care about. The other columns describe the customer as they were before they left or stayed, for example:

• how long they have been a customer (tenure);
• what they buy and how much they pay each month;
• contract or plan type, and how they pay;
• recent usage, support contacts or complaints — if recorded before the outcome.

Leave out anything recorded because the customer left: cancellation reasons, final-bill flags, "sent to win-back" markers. Those columns give the answer away in historical data and are empty when you need a prediction. This is target leakage, and it is the most common reason churn models look excellent in testing and fail in use. Customer IDs should go too: they let a model memorise records instead of learning patterns.

A measured example: 7,043 telecom customers

What we measured

• Data: the IBM Telco Customer Churn sample — 7,043 customers, of whom 26.5% churned. Each row has the customer's tenure, services, contract type, payment method and charges.
• Test design: 20% of customers (1,408) were held back before any model was trained; every result below comes from them.
• Model: YourCloudGroup Model Manager ran its twelve-variant model tournament; a random forest won, with a ranking quality (AUC) of 0.83.
• Leakage check: the customer ID was excluded automatically; no other column was flagged.
• Placeholder costs: a retention offer costs $60; it keeps 30% of customers who would otherwise leave; a kept customer is worth 12 months of their bill.

How much each strategy earned on the held-back customers

What the result means

The value peaks when roughly the top half of the ranked list gets the offer. Below that point, each extra offer goes to someone with a real chance of leaving; beyond it, offers increasingly go to customers who were staying anyway, and the $60 cost outweighs the saving. Offering everyone still makes money in this example, but less than a third of what the ranking makes.

What the result does not show

The costs are placeholders: with a cheaper offer, the best cut-off moves further down the list; with a more expensive one, it moves up. The data is a public sample, not your customers. The top 20% of the list churned at 2.4 times the average rate — just short of the 2.5× target set before the test — so the model is useful but not spectacular. And because the dataset has no dates, the timing of each customer's departure could not be checked; the certificate records that as "not assessed" rather than assuming it.

How to predict churn in practice

1. Export one row per customer with a left/stayed column for a fixed period (for example, "cancelled within 90 days").
2. Remove post-churn columns and identifiers, and have someone who knows the data confirm the list.
3. Train on most of the history and test on a held-back part — never judge a model on customers it learned from (held-out evaluation).
4. Judge the ranking, not the accuracy. With 25% churn, a model that says "nobody leaves" is 75% accurate and useless; look at how many leavers sit at the top of the list, or at balanced accuracy.
5. Put your own costs on it and pick the cut-off where value peaks.
6. Score current customers each month and act on the top of the list; re-check the model as new outcomes arrive.

How YourCloudGroup Model Manager does it

YourCloudGroup Model Manager runs these steps from a CSV or spreadsheet in a browser, on a dedicated server: it seals a held-back test set, screens columns for leakage and identifiers, compares twelve model variants on the same test, and issues a certificate stating what was proven and what was not. Trained models can score new customer lists on a schedule. See the full set of business trials for the method and the other two cases.

Takeaway

Churn prediction is a ranking problem: the question is not "will this customer leave?" but "who should we talk to first?". With a few thousand past customers and an honest test, a model can concentrate leavers at the top of the list — in this example, almost three times the average rate — and your own costs decide how far down the list to go.

Sources

• IBM, Telco Customer Churn sample dataset.
• YourCloudGroup blind-set trials, 30 September 2026: Predictions from business spreadsheets.