# Churn rate: what it is, how to calculate it, how to analyze it

Page: https://alidemirbas.com.tr/en/blog/churn-rate-calculation-analysis

2026-11-08 · CRM & Lifecycle · Ali Demirbaş

Churn rate is the share of customers lost in a period. Three denominators, how to annualize a monthly rate, inactivity windows and how to analyze churn.

Churn rate is the share of customers who stop using a product or service during a given period. In the basic calculation, you count how many of the customers you had at the start of the period were lost during it. You divide that by the starting count and multiply by 100.

Even one month of data can produce several churn rates. In the hypothetical example below, the same month gives 6.0% or 7.5% churn depending on the denominator. Three choices decide the result: when a customer counts as lost, who goes into the denominator, and how the rate is converted to another period.

The [guide to retention, churn and lifetime value](https://alidemirbas.com.tr/en/blog/customer-retention-metrics-churn-ltv) covers how the three relate. The start-of-period calculation runs in the [retention rate calculator](https://alidemirbas.com.tr/en/calculators/retention-rate). Its new-customer input takes only new customers still active at period end, not everyone who joined.

## What churn means in business

[Oxford Learner's Dictionaries](https://www.oxfordlearnersdictionaries.com/definition/english/churn_2) lists the business sense of churn alongside "churn rate". It means the number of people who stop using a product and switch to another one. The noun's first sense is a butter churn. The dictionary also uses the business sense for employees who leave, while in marketing and customer relationship management (CRM) it means customer loss.

People use churn for two different things. One is the event of a single customer leaving. The other is the rate of loss across a period. Keeping them apart helps whenever a dashboard says "churn" without saying which one it means.

## When a customer counts as churned depends on the relationship

Whether churn is observed at all depends on whether the relationship is contractual. Peter Fader and Bruce Hardie draw this line in a 2007 [paper on projecting customer retention](https://faculty.wharton.upenn.edu/wp-content/uploads/2012/04/Fader_hardie_jim_07.pdf). In a contractual setting, the firm sees when a customer becomes inactive, for example when a subscription isn't renewed. In a noncontractual setting, customers don't announce that they've left, they just stop coming back.

The same authors' 2009 [overview of probability models for customer-base analysis](https://faculty.wharton.upenn.edu/wp-content/uploads/2012/04/Fader_hardie_jim_09.pdf) names the core problem of noncontractual settings. You have to tell a customer who has ended the relationship from one who is only in a long pause between purchases.

Their example is Amazon's definition of an active customer as one who ordered in the past year. The authors call that 12-month cut-off arbitrary. Move it to nine months and the apparent customer base shrinks, while the true base stays the same.

The two settings lead to two kinds of churn definition. With subscriptions, memberships or phone contracts, churn is an event with a date. In ecommerce or an app without a subscription, churn follows from an inactivity period you choose. When that period moves, the rate moves too.

## How to calculate churn rate: three denominators for the same month

The numerator and denominator of the churn rate formula change with who gets counted. Three versions need to be kept apart:

- Start-of-period base: Customers lost from those who started the period, divided by the customers at the start. Customers who join and leave within the same period aren't counted. The site's calculator uses this method.
- Period average: All customers lost during the period, divided by the average number of customers in the period.
- Cohort base: Customers who joined in the same period are followed as a fixed group. You measure the share lost in their first month, first three months or first year.

Fader and Hardie's 2010 [paper on valuing a contractual customer base](http://brucehardie.com/papers/022/fader_hardie_mksc_10.pdf) cites a PricewaterhouseCoopers report on setting the number at risk. Its two standard ways are the customer count at the start of the period and the average of the opening and closing counts.

A public example of the second method is the churn definition in T-Mobile's [Form 10-K for fiscal year 2023](https://investor.t-mobile.com/static-files/6c8f67e6-5485-457f-b57e-cb100ce39e34). There, churn is deactivated customers as a percentage of the average number of customers in the period. The result is "further divided by the number of months in the period".

Deactivations are counted net of customers whose service was restored within a certain period, and customers served for less than a minimum period are excluded. In a quarterly report, dividing by three turns the quarter's rate into a monthly average.

A cohort rate isn't directly comparable with the other two, because it depends on how long customers have been with you. In the same business, customers in their first month and two-year customers can leave at different speeds.

### Hypothetical example: four rates from one month of data

Take a hypothetical subscription business that starts the month with 2,000 subscribers, 120 of whom cancel during the month.

In the same month, 400 new subscribers join and 40 of them leave before the month ends. That leaves 2,240 subscribers at month end. The table shows how four calculations turn this data into four different rates.

*Hypothetical example: four churn rates from one month of data. The average base is an approximation, the mean of the start and end counts. Rates are rounded to one decimal place.*

| Method | Numerator | Denominator | Churn rate |
| --- | --- | --- | --- |
| Start-of-period base | 120 | 2,000 | 6.0% |
| Period average (approximate) | 160 | 2,120 | 7.5% |
| Joined during the month, to month end | 40 | 400 | 10.0% |
| Mixed calculation | 160 | 2,000 | 8.0% |

The first three rows are correct, but they answer different questions. The start-of-period base shows how much of the existing base left, and the period average shows gross loss against the average base. The base grew from 2,000 to 2,240 this month, so the second row is not a shrink rate. The fourth row mixes bases: it adds new subscribers' losses to the numerator but leaves them out of the denominator, which inflates the rate.

The third row shows losses among subscribers who joined during the month, up to month end. They joined on different days, so each was observed for a different length of time, and the row doesn't measure each subscriber's first 30 days. For first-30-day churn, start the clock at each subscriber's join date and keep those with fewer than 30 days aside.

The gap between methods grows with the number of new subscribers in the month. In the next table, the starting base stays at 2,000 and its monthly loss at 6%. The loss among subscribers who joined during the month stays at 10% up to month end. Only the number of new subscribers changes.

*Hypothetical example: three methods by number of new subscribers, with the same rounding as above.*

| New subscribers in the month | Start-of-period base | Period average | Mixed calculation |
| --- | --- | --- | --- |
| 0 | 6.0% | 6.2% | 6.0% |
| 400 | 6.0% | 7.5% | 8.0% |
| 1,000 | 6.0% | 9.2% | 11.0% |

Even with no new subscribers, the period average comes out slightly higher, because a base that shrinks during the month lowers the denominator. In a fast-growing business, the same customer behavior can produce rates from 6% to 11% just by changing the method. That's why a reported churn rate needs the name of its base next to it.

## Monthly churn doesn't convert to annual churn by multiplying by 12

Converting a monthly churn rate into an annual one has to account for each month's loss acting on the customers left from the month before. Assume the same rate applies each month to the customers active at its start, with no new customers and no returns. Annual churn is then 1 − (1 − monthly churn)¹².

The formula is built for start-of-period monthly rates on one group, not for the share inactive under a window. A measured annual churn follows the customers at the start of the year directly to year end.

In a hypothetical example, monthly churn is 3%. Multiplying by 12 gives 36%, while the compounded calculation gives about 30.6%.

Under the same assumption, a base of 1,000 customers that adds no new customers falls to about 694 by year end. Estimating average customer lifetime as 1 divided by monthly churn rests on the same constant-rate assumption, and here it gives about 33 months.

The constant-rate assumption often fails for real cohorts. Fader and Hardie's 2010 paper notes that retention rates typically increase over time at the cohort level.

The rate here is period-to-period. In the 2007 paper's definition, it's the share of customers active at one period's end who are still active at the next period's end. The share left of the starting group still declines.

The 2010 paper adds that the increase has been attributed, in large part, to a sorting effect in a heterogeneous population. The 2007 paper describes the mechanism within its model: high-churn customers drop out early, and those who remain have lower churn probabilities. In that model, individual churn probabilities don't change over time, only the mix of who remains.

The same paper also shows that a calculation built on a single retention rate underestimates the residual value of the existing customer base. One inference follows: spreading a new cohort's first-month churn across twelve months can overstate that cohort's losses by year end. To calculate lifetime value itself, you can use the [customer lifetime value (LTV) calculator](https://alidemirbas.com.tr/en/calculators/ltv). Note which cohort and which method your churn input comes from, so the result reads correctly.

## Without a contract, an inactivity window defines churn

In a noncontractual business, the churn rate depends on how many days of inactivity make a customer count as lost. A short window counts regular customers who are only taking a long break. A long window spots losses late, and any win-back attempt starts late too.

One measurable way to choose the window is to test it on past data. Pick a past analysis date and find who each window would have counted as churned. Then check how many of them ordered again in the following months.

Some of them were counted as churned by mistake, while others may have left and then come back. The data can't tell the two apart, it only shows the reorder share.

Take a hypothetical ecommerce business with 5,000 customers who ordered at least once in the past 12 months. The table shows the customers inactive at the analysis date under three windows. It also shows how many of them ordered again within the next 180 days.

*Hypothetical example: three windows on the same customer base. The percentages in the second column describe one date, not a monthly churn rate. They are whole numbers, so no rounding applies.*

| Inactivity window | Inactive at analysis date | Ordered within 180 days | Reorder share |
| --- | --- | --- | --- |
| 60 days | 1,400 (28%) | 420 | 30% |
| 90 days | 1,000 (20%) | 200 | 20% |
| 180 days | 600 (12%) | 60 | 10% |

As the window gets longer, both the inactive share and the reorder share fall. On the other hand, the 180-day window shows a loss only after six months. The choice balances counting customers who will come back against detecting losses late. For a monthly churn rate, count the customers who cross the threshold for the first time that month.

Whatever window you pick, write it next to the report and keep it fixed across periods. As the Amazon example shows, a new window changes the rate even when the customers stay the same.

Time since last order is the first dimension of [RFM analysis](https://alidemirbas.com.tr/en/blog/rfm-analysis-customer-segmentation), which segments customers by recency, frequency and monetary value (RFM). The flow for reaching customers past the window lives in the [lapsed customer win-back journey](https://alidemirbas.com.tr/en/lab/journeys/lapsed-customer-win-back). Instead of a fixed window, probability models estimate each customer's chance of still being active, and they belong to the churn analysis toolkit.

## Customer churn and revenue churn measure different losses

Churn measured in customers shows how many customers left, while revenue churn shows how much recurring revenue was lost. [ChartMogul's revenue churn page](https://chartmogul.com/saas-metrics/revenue-churn/) defines gross revenue churn as monthly recurring revenue (MRR) lost to cancellations and downgrades. It divides that loss by MRR at the start of the period.

Net revenue churn subtracts revenue gained from expansion and reactivation. Both rates use the same denominator, MRR at the start of the period, and revenue from new customers doesn't appear in either formula. That's why net revenue churn can go negative while gross revenue churn can't.

The two measures support different decisions. Customer churn tracks how many people the product keeps, and revenue churn tracks how well the business protects its revenue base. A few large accounts leaving can look small in customer churn and large in revenue churn. To track the revenue side, you can use the [net revenue retention (NRR) calculator](https://alidemirbas.com.tr/en/calculators/nrr).

### Voluntary and involuntary churn

The reason a customer leaves is a split of its own, separate from the unit of measurement. Voluntary churn is a customer's own decision to cancel. Involuntary churn is a subscription that ends because a payment failed, even though the customer didn't mean to leave.

Summed into one rate, a product problem and a payment-flow problem can't be told apart. The steps for a subscriber whose payment failed are in the [payment failure recovery journey](https://alidemirbas.com.tr/en/lab/journeys/payment-failure-recovery).

## How to analyze churn: cohorts, segments, reasons and risk

Churn analysis breaks losses apart to find which customers, which periods and which causes move the churn rate. A workable order is cohorts first, then segments, then reasons, and prediction last, once the definition and the window are fixed.

### Cohort survival curves

A cohort survival curve shows how much of a starting group is left at the end of each month. The [Kaplan-Meier section of the NIST statistics handbook](https://www.itl.nist.gov/div898/handbook/apr/section2/apr215.htm) explains that the method works with complete or censored samples and assumes no particular distribution. A customer is censored when observation stops while they're still active, so their eventual churn date is unknown.

Pooling cohorts without accounting for censoring biases the curve. A cohort that joined last month can't have reached month six yet, so dividing month-six losses by every customer understates churn. A steep early drop that later flattens is consistent with the sorting effect described above. It points to the first weeks as the place to look first.

### Segments

Segment comparisons split churn by plan, acquisition channel, region or customer value. Two checks keep them reliable. A small segment's rate swings from month to month for reasons unrelated to the product, so compare counts as well as rates. The total rate can also change when the segment mix changes, even if no segment's own rate moves.

For noncontractual businesses, RFM segments give a ready-made split by recency, frequency and spend. Choose the segments before you look at the results, so the comparison doesn't turn into a search for whichever slice looks worst.

### Reasons

Reason analysis looks at why customers leave. Cancellation surveys only reach customers who cancel actively, so they miss involuntary churn and silent lapsing. Usage before churn, such as falling activity or unresolved support tickets, points to candidate causes. A signal like this is a lead to test, not proof, until an experiment or the raw records confirm it.

### Prediction and targeting

Churn prediction scores each customer's risk of leaving within a set window. Eva Ascarza's [Journal of Marketing Research paper](https://www.hbs.edu/ris/Publication%20Files/ascarza_jmr_18_783d54d4-e548-41ed-b1d7-8a180f1ae85a.pdf) is based on two field experiments. One ran at a wireless provider in the Middle East, the other at a membership organization in North America. It found that customers with the highest churn risk aren't necessarily the right targets for a retention campaign.

In the wireless study, 52% of the top half by risk were also in the top half by response. In the membership study, the figure was 40%. Random selection would give 50%, so the second study's overlap was below random. Overall, the author finds the relationship no different from independence.

The practical step is to run a retention campaign with a randomly held-out control group. The control group gives the campaign's average effect. Targeting by response needs a model of who responded, checked on a separate sample, as the paper does with a random calibration and validation split.

How to escalate an intervention once a risk threshold is crossed is set out in the [tiered churn intervention journey](https://alidemirbas.com.tr/en/lab/journeys/churn-risk-escalation). What vendor churn scores can and can't tell you is covered in the [post on AI in lifecycle marketing](https://alidemirbas.com.tr/en/blog/ai-in-lifecycle-marketing).

## Five questions before comparing two churn rates

Two churn rates side by side mean something only if both count the same thing. Before you compare your rate with one from another company's report, a tool's dashboard or last year's deck, ask these five questions:

1. How was a loss defined: by cancellation date, by the end of the paid period, or by an inactivity window?
2. Is the denominator the customers at the start, the period average, or a fixed cohort?
3. Is the rate monthly or annual? Was the annual rate found by multiplying by 12?
4. Who was left out: trial users, customers below a minimum service period, customers who returned within a set period? T-Mobile's definition doesn't count the last two as losses.
5. Does the rate count customers or revenue? Is involuntary churn included?

If these answers aren't written down, the gap between two rates may come from the method rather than from customer behavior. Writing them once under every report keeps comparisons sound months later.