Churn risk analysis

Churn risk analysis is the practice of scoring how likely each customer is to cancel within a defined window, then ranking accounts by that score so the team can act on the riskiest ones first. Instead of measuring churn after it has already happened, it estimates churn before it happens. A customer whose product usage has dropped 60 percent over two months, who has logged three support escalations, and whose champion has left the company carries a high risk score. A customer using the product daily with a renewal six months out and a recent expansion carries a low one.

The value of the analysis is in the ranking, not the raw number. If you have 400 accounts and a customer success team that can have meaningful conversations with 30 of them this month, the question is not how many will churn but which 30 to call. Churn risk analysis answers that. It concentrates limited attention on the accounts where intervention is both needed and still possible.

A good risk model combines several signal categories. Engagement signals cover how the product is being used. Relationship signals cover the health of the human connection, such as champion changes and sentiment in support tickets. Commercial signals cover money, such as failed payments, downgrades, and contract value relative to seats actually used. No single signal is reliable on its own. A drop in usage might mean a seasonal lull or it might mean the account is winding down, and only the combination of signals separates the two.

A churn risk score is built by selecting the signals that predict cancellation, weighting each one by how strongly it has predicted churn in the past, summing the weighted signals for each account, and normalising the result to a comparable scale such as 0 to 100. The weighting is what separates a useful model from a noisy one. You set weights by looking at customers who churned and finding which signals consistently appeared before they left.

For example, suppose monthly active usage falling below 30 percent of licensed seats carries a weight of 40, a failed payment in the last 30 days carries 25, and a departed champion carries 20. An account showing all three would score 85 out of a maximum of 85, placing it in the highest risk band. An account showing only the usage drop would score 40, a moderate risk worth watching but not yet urgent. The same model applied to every account produces a ranked list the team can work top down.

1. 1

   Select the risk signals

   Choose indicators that have a plausible causal link to cancellation: declining active usage, reduced login frequency, failed payments, support escalations, low feature adoption, and changes in the buying contact. Start with signals you can measure reliably.

2. 2

   Weight each signal by predictive strength

   Look at customers who churned and measure how often each signal appeared beforehand. Signals that reliably precede churn get higher weights. Signals that appear equally in retained and churned accounts get low weights or are dropped.

3. 3

   Score every account on each signal

   For each customer, record which signals are currently present and at what intensity. A 60 percent usage drop should contribute more than a 10 percent drop, so use graded thresholds rather than simple yes or no flags where you can.

4. 4

   Sum and normalise to a comparable scale

   Add the weighted signals per account and divide by the maximum possible score to get a 0 to 100 figure. Group scores into bands such as low, medium, and high so the team has a clear queue to work through.

A churn risk score is a composite, which makes it a natural fit for a metric tree. The score sits at the top and the signal categories that feed it sit beneath, each broken into the specific measures that drive it. Modelling it this way stops the score from being a black box. When an account moves into the high risk band, the tree shows which branch pushed it there.

There is no universal benchmark for a churn risk score because the scale and weights are specific to each business. What you can benchmark is how the score is distributed and how well it performs. A healthy model concentrates real churn in the top risk band and rarely lets a churned account come from the low band. The ranges below describe what good distribution and accuracy look like for a subscription business.

Judge the model against its own history rather than an external number. Track how many high risk accounts actually churned each quarter and refine the weights when the gap widens. A model that captured 70 percent of churn last year but only 45 percent this year is drifting, usually because customer behaviour has changed and the old signals no longer carry the same meaning.

Improving churn risk analysis means making the score more accurate and making the team faster to act on it. Accuracy comes from better signals and honest calibration against what actually happened. Speed comes from routing each flagged account to the person who can do something about it, with enough context to act.

1. 1

   Treating the score as the whole story

   A risk score points the team at an account, it does not explain it. Teams that act on the number without talking to the customer often address the wrong cause and lose the account anyway.

2. 2

   Flagging too many accounts

   If the high risk band holds a third of the base, the team cannot work it and starts ignoring the list. A model is only useful if its top band is small enough to action.

3. 3

   Never recalibrating the weights

   Signals decay. A behaviour that predicted churn two years ago may be normal now. A model that is never compared against actual outcomes slowly stops working while still looking authoritative.

4. 4

   Measuring risk without measuring action

   Scoring risk and then doing nothing differently changes no outcomes. The point of the analysis is to drive intervention, so track whether flagged accounts were actually contacted and whether it helped.