Why Churn Is Harder to Model Than It Looks

Growth suppresses reported churn, cohorts fix that and introduce censoring, and half of it may be failed payments. There is no single churn number, and the search for one is why the subject stays confusing.

KISSmetrics Editorial

|11 min read

Churn is hard to model because the obvious formula quietly changes meaning whenever your growth rate changes. Cancellations divided by customers is not a property of your retention; it is a property of your retention and your recent acquisition, mixed together. A month of strong signups lowers churn without anyone retaining anyone.

This is why the standard number misleads, why the cohort fix is incomplete, and what to report instead of pretending there is one churn rate.

The newsletter

Join our KISS newsletter

One short read a week on what actually moves revenue, in a free email. Read by 10,000+ operators and founders.

No spam. Unsubscribe in one click.

I.The denominator problem, which is most of it

Cancellations over customers mixes retention with acquisition, so the number moves for reasons that have nothing to do with retention.

A.Why growth flatters you

Take 1,000 customers, 50 cancellations, 5% churn. Next month you add 500 new customers and still lose 50. Churn is now 3.3%.

Nothing about retention improved. The denominator grew with people who have not had time to leave yet, and the reported figure fell. The reverse is worse: a company whose growth slows watches churn climb and concludes the product got worse, when the arithmetic simply stopped hiding it.

Any company whose growth rate is changing has a churn number that is moving for reasons unrelated to churn.

B.Customer churn and revenue churn disagreeing, correctly

Lose fifty small accounts and keep two large ones and customer churn is alarming while revenue churn is fine. Lose one enterprise account and a hundred starters stay, and the reverse.

Neither is the real number. They answer different questions: whether you are keeping people, and whether you are keeping money. A business reporting only one of them is blind on that side, and net revenue retention hides both by netting expansion against loss. See net revenue retention.

C.Voluntary and involuntary are not the same event

A customer who cancels has made a decision about your product. A customer whose card expired has not made any decision at all.

Involuntary churn is frequently a large share of the total and it is a billing problem: retries, card updater, dunning email timing. Mixing the two produces a retention programme aimed at people who never intended to leave, and hides a payments fix that is usually cheaper than anything the product team could do.

Split them before anything else. It is the highest-value ten minutes in this whole subject.

II.The cohorts that have not finished churning

Cohorts fix the denominator and introduce a subtler error that reads as improvement.

A.What cohorting fixes

Group customers by when they arrived and follow each group forward. Now the denominator is fixed: the January cohort is always the same people, so a change in its retention is genuinely a change in retention.

This is the right base for everything below, and it is why cohort retention rather than monthly churn is the number worth reporting.

B.And the error it introduces

Right-censoring. The January cohort has had eight months to churn. The July cohort has had one. Comparing their retention rates as though both were finished makes recent cohorts look better than they are.

This produces a specific and common false conclusion: a chart where recent cohorts appear to be improving, presented as evidence that a product change worked, when it is the shape of an incomplete cohort every time.

The fix is to compare at fixed ages. Retention at day 30 for every cohort, at day 90 for every cohort old enough to have one. Never compare a three-month-old cohort's lifetime retention to a two-year-old cohort's.

Retention by cohort, and the triangle that misleads

Cohorts view
CohortM1M2M3M4M5M6
Jan
88%
79%
74%
71%
69%
68%
Feb
89%
80%
75%
72%
70%
Mar
90%
82%
77%
74%
Apr
91%
83%
78%
May
92%
84%
Jun
93%
Illustrative, not measured. Read down a column and retention is genuinely improving. Average across each row, which is what a summary does, and June looks spectacular purely because it only contains M1.

III.What to do instead of solving it

There is no single correct churn number. Report the small set that together is honest.

A.The four numbers to report

Retention at fixed ages, by cohort. Day 30, 90, 180, 365. Comparable across cohorts, immune to the growth-rate problem, and censored only in the sense that a young cohort simply has no figure yet rather than a flattering one.

Voluntary and involuntary, separately. Always. They go to different teams.

Customer and revenue, separately. Not netted.

The survival curve, not the average lifetime. Average customer lifetime is dominated by a long tail and is close to meaningless in most subscription businesses. The shape tells you whether you have an onboarding problem or a long-run value problem, which the average cannot.

Why each common single number fails on its own

Diagnosis view
NumberFails whenReport instead
Monthly churn %Growth rate changesCohort retention at fixed ages
Cohort lifetime retentionCohorts differ in ageThe same fixed age for all
Net revenue retentionExpansion masks lossGross and net, side by side
Average customer lifetimeThe tail is longThe survival curve
Total churnPayments are failingVoluntary and involuntary split
No row is wrong, and every row is incomplete alone. The set is the report.

B.Prediction, and why it is harder than it looks

Predicting which accounts will churn is the natural next step and it inherits every problem above plus one more: the outcome you are training on is censored. Accounts that have not churned yet are not the same as accounts that will not.

The pragmatic version most teams should run instead is not a model. It is a behavioural threshold, derived from comparing what churned accounts did in the weeks before cancelling against what retained ones did. Less sophisticated, considerably more likely to be acted on. Our churn prevention workflow covers wiring that up, and diagnosing churn covers finding the signal.

C.What makes the honest version practical

All of this needs per-person history with the payments attached: when each customer arrived, what they did month by month, when they stopped, and whether they chose to.

Kissmetrics groups people by when they first appeared and follows them forward on the same record that carries their behaviour, so retention at a fixed age is a report rather than an export, and any cell in the grid opens into the accounts inside it. That last part is what turns a retention number into something a customer success team can act on this week.

It does not resolve the censoring. Nothing does except waiting, and the discipline of comparing at fixed ages is a human habit rather than a feature.

Verdict

There is no single churn number, and the search for one is why the subject stays confusing. Monthly churn moves with your growth rate, cohort retention is censored for recent cohorts, customer and revenue churn disagree legitimately, and half of it may be failed payments.

Split voluntary from involuntary first, then report cohort retention at fixed ages with customer and revenue separately. Compare a cohort only against cohorts of the same age. A young cohort with no figure is more honest than a young cohort with a flattering one.

Continue Reading

churn ratechurn modelingcohort retentioninvoluntary churnretention analysissurvival curve
KISSmetrics

Build your business intelligence layer for free.

KISSmetrics groups people by when they first appeared and follows them forward on the record that carries their behaviour, so any cell in the grid opens into the accounts inside it.