Cohort analysis: retention table, curve, LTV forecast and RFM as mathematics
How to build a cohort retention table, fit a curve, extend the tail and obtain an LTV forecast, and also how to compute RFM segments by quantiles and gauge revenue concentration. The decision: which cohorts got better or worse, what a customer is worth and which segments to target with communications.
Why this document
The head of CRM and marketing read this to tell whether customers are getting better or worse and what they will bring over a year or two. The overall average retention hides a change in the quality of new customers, so customers are grouped by month of first purchase and cohorts are compared at the same age. A retention curve lets you extend the tail and estimate LTV before data accumulates, and RFM shows who brings most revenue. Without cohorts, a channel that brought one-time buyers looks successful on first-month revenue.
At a glance
- What is this document and why is it needed?
- How to build a cohort retention table, fit a curve, extend the tail and obtain an LTV forecast, and also how to compute RFM segments by quantiles and gauge revenue concentration. The decision: which cohorts got better or worse, what a customer is worth and which segments to target with communications.
- Who owns the document and how often is it rebuilt?
- CRM and loyalty analyst; the cohort table is updated monthly, RFM is recomputed monthly or weekly before mailings; read by the head of marketing, e-commerce and the finance lead.
- What mistakes are the most common?
- A simple average over cohorts is used → small cohorts distort the curve → weight by size. Cohorts of different channels are mixed → good CRM retention masks poor paid channels → slice by channel. The curve is extrapolated far beyond the observed horizon → LTV rests on an assumption → show observed and forecast parts separately.
For whom
- Brand
- A brand builds cohorts for D2C and the loyalty program; marketplace customers belong to the platform, so cohorts for them are built by brand (repeat brand purchases) if the platform provides data. The main aim is an LTV forecast for the CAC decision.
- Retail
- A retailer merges stores and online by loyalty card: a cohort is customers with a first purchase in a month and retention is counted by any purchase in the chain. A large share of anonymous receipts understates retention, and this is stated in the report.
What it consists of
The cohort retention table
- A table row is a cohort (month of first purchase), a column is age in months, the value is the share of the cohort that bought in that month of life. The first column (age 0) is always 100%; the lower right part is empty because young cohorts have not lived those months yet.
- In the example there are six cohorts (1,250, 1,480, 1,120, 1,630, 1,390, 1,710 customers) observed for 12 months. Weighted retention is 27.5% in month 1, 16.8% in month 3, 11.9% in month 6, 9.9% in month 11.
- The table is read by rows and diagonals: a row shows cohort quality, a column compares cohorts at the same age, a diagonal reveals the calendar effect (season, promotions, catalog change) on all cohorts at once.
Retention curve and LTV forecast
- Average retention is well described by a power curve r(t) = a·t^(−b): regression of ln r on ln t over ages 1-11 gives a = 0.27, b = 0.43, R² = 0.99 (true values are set in the model). The curve falls fast in the first months and slowly later; an exponential model overstates tail losses.
- LTV = CM2 per order × Σ r(t)/(1 + δ)^t at CM2 RUB 1,350 and δ 1.5% a month: RUB 3,452 over 12 months, RUB 4,429 over 24 months (curve forecast); the observed part over 11 months is RUB 3,347. The tail (months after observation) is an estimate and is shown in a separate color.
- Brand: D2C LTV is compared with CAC and with marketplace margin; Retailer: chain customer LTV includes store and online purchases and the CAC of a store opening is spread over the zone customers.
RFM as mathematics
- R (recency) is days since the last purchase, F (frequency) the number of orders, M (monetary) the spend; each scale is split into quintiles and scored 1-5 (for R inverted: fresher is higher). A segment is set by a rule on scores, not by eye.
- In the example of 600 customers each segment has two different shares: the share of customers and the share of revenue. Champions 66 (11% of customers, 25% of revenue), loyal 94 (15.7% of customers, 28.5% of revenue), at risk 56 (9.3% of customers, 9% of revenue), lost 130 (21.7% of customers, 9.9% of revenue). The top 20% of customers give 48.5% of revenue, the top 10% 31.7%; this is concentration by customer, not by segment.
- A segment works as an action: champions get the loyalty program without discounts, the at-risk segment a personal win-back, the lost a limited reactivation; communication cost is checked against segment contribution.
Brand and retailer: cohorts and segments in practice
- Brand: brand cohorts are built for D2C and the loyalty program; on a marketplace and in wholesale a customer cohort cannot be built, so repeat brand purchases at the partner or from anonymized platform data are used.
- Retailer: retailer cohorts are built by loyalty card across the chain; the share of anonymous receipts understates retention, so it is estimated and shown in the report with the table.
- Brand: a brand uses the retention curve for D2C LTV and the permissible CAC decision; comparing cohorts by acquisition channel shows which channels bring loyal customers.
- Retailer: a retailer uses the retention curve to value a store-zone customer and for the opening decision and the loyalty budget; cohorts are split by format and region.
- Brand: RFM for a brand is built on D2C customers: champions get early access and collaborations, the at-risk segment a personal win-back; purchase frequency at a premium brand is lower and F quintiles are chosen carefully.
- Retailer: RFM for a retailer is built on chain customers: frequency is higher, and RFM relies on categories and stores; segments are tied to the mailing schedule and the loyalty program benefits.
- Brand: a brand checks revenue concentration by customer separately for D2C and by partner: dependence on one partner is a risk and not an RFM segment.
- Retailer: a retailer checks revenue concentration by customer by store zone: a high share of a small group of customers makes the store vulnerable to their leaving.
- Brand: a brand measures the effect of communications by segment on a D2C control group; for partners the effect of brand activity is assessed by sell-out against control points.
- Retailer: a retailer measures the effect of communications by segment on card control groups and control stores; personal offers must not conflict with the general pricing policy.
Where the data comes from
| Data source | Fields |
|---|---|
| Receipts with customer ID | Date, amount, channel, customer identifier (card, phone, account) with merge rules; the share of anonymous receipts separately. |
| Cohort register | Month of first purchase, acquisition channel, region, first category; flags for cohort slicing. |
| Order contribution | CM2 per order and by cohort from the unit economics document, discount rate. |
How to build it step by step
- Step 1. Define “customer” and “activity”: by card or account, any chain purchase or only the brand; write the rule down.
- Brand: Step 2. Separate D2C, marketplaces and wholesale; for marketplaces use platform data on repeat brand purchases.
- Retailer: Step 3. Merge stores and online by card and estimate the share of anonymous receipts that understates retention.
- Step 4. Assign each customer a month of first purchase and build the “cohort × age” table with shares.
- Step 5. Compute the weighted curve by cohort size and compare cohorts at the same age; mark diagonal effects.
- Step 6. Fit a curve (power or other) on the observed part and check residuals; do not extrapolate beyond twice the observed horizon without a caveat.
- Step 7. Compute LTV on CM2 with discounting for 12 and 24 months and separate the observed and forecast parts.
- Step 8. Compute R, F, M, quintile scores and the segment rule; check segment sizes and revenue share.
- Step 9. Fix actions per segment and measure their effect on a control group (see the experiments document).
- Step 10. Publish the table with the number of customers per cohort, an interval and the update date; mark measures for small cohorts as unstable.
Formulas
| Measure | Formulas | Note |
|---|---|---|
| Cohort retention | r_(k,t) = customers of cohort k who bought in life month t / size of cohort k | r_(k,0) = 1; customers are counted by unique identifiers. |
| Weighted curve | r̄_t = Σ_k n_k·r_(k,t) / Σ_k n_k over cohorts observing age t | A simple average over cohorts is skewed by small cohorts. |
| Power curve | r(t) = a·t^(−b); ln r = ln a − b·ln t | OLS fit on logs, ages t ≥ 1; the curve is undefined at t = 0. |
| LTV from the curve | LTV_H = cm2·Σ_(t=0..H) r(t)/(1 + δ)^t, r(0) = 1 | The observed part comes from the table, the rest from the curve. |
| Quintile score | s(x) = 1 + Σ_(j=1..4) 1[x > Q_(0.2·j)]; for R: 6 − s(R) | Quintiles over period customers; with tied values scores are uneven. |
| Revenue concentration | C_20 = Σ_(i ≤ 0.2·N) M_(i) / Σ M_i, M_(i) in descending order | The same device as Pareto by SKU. |
Formulas in notation
- Retention
r_{k,t} = N_{k,t} / N_{k,0}
N_{k,t} customers of cohort k who bought in month t.- LTV
LTV_H = cm2 · Σ_{t=0..H} r(t) / (1 + δ)^t
Discounted expected customer contribution.- Concentration
C_p = Σ_{i ≤ pN} M_(i) / Σ_i M_i
Revenue share of the top p customers.
Worked example
Retention of six cohorts by life month, % (empty where the cohort has not lived the month yet)
| Cohort | Customers | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | M10 | M11 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Jan | 1,250 | 32.7 | 22.4 | 16.2 | 16.2 | 15.5 | 12.8 | 12.1 | 11.8 | 11.9 | 10.9 | 9.9 |
| Feb | 1,480 | 27.8 | 19.3 | 18.1 | 16.8 | 16 | 12.9 | 12.4 | 10.1 | 9.5 | 9.7 | |
| Mar | 1,120 | 26 | 22.4 | 17.8 | 16.5 | 13.2 | 12.3 | 11.4 | 10.5 | 11.6 | ||
| Apr | 1,630 | 28 | 19.2 | 17.1 | 14.8 | 13.7 | 12.5 | 11.5 | 11.3 | |||
| May | 1,390 | 26.7 | 21.5 | 14.6 | 14.4 | 13.3 | 11.5 | 12.1 | ||||
| Jun | 1,710 | 24.5 | 17.3 | 16.9 | 13.5 | 11.7 | 9.8 | |||||
| Weighted average | 8,580 | 27.5 | 20.1 | 16.8 | 15.2 | 13.8 | 11.9 | 11.9 | 10.9 | 10.9 | 10.3 | 9.9 |
The table is generated in code: a true curve of 0.30·t^−0.45, a 3% quality decline per newer cohort and binomial noise. The fitted curve has a = 0.27, b = 0.43; 24-month LTV is RUB 4,429. The data is illustrative.
Figures
- 9.8 … 32.7 %
Data as a table
| M1 | M2 | M3 | M4 | M5 | M6 | |
|---|---|---|---|---|---|---|
| Jan | 32.7 | 22.4 | 16.2 | 16.2 | 15.5 | 12.8 |
| Feb | 27.8 | 19.3 | 18.1 | 16.8 | 16 | 12.9 |
| Mar | 26 | 22.4 | 17.8 | 16.5 | 13.2 | 12.3 |
| Apr | 28 | 19.2 | 17.1 | 14.8 | 13.7 | 12.5 |
| May | 26.7 | 21.5 | 14.6 | 14.4 | 13.3 | 11.5 |
| Jun | 24.5 | 17.3 | 16.9 | 13.5 | 11.7 | 9.8 |
Cohort retention by life month. Rows are cohorts, columns are ages 1-6 months (share of customers who bought in the month); color shows the level.
- Observed, %
- Fit, %
Data as a table
| Measure | Observed, % | Fit, % |
|---|---|---|
| 1 | 27.5 | 27.1 |
| 2 | 20.1 | 20.2 |
| 3 | 16.8 | 17 |
| 4 | 15.2 | 15 |
| 5 | 13.8 | 13.7 |
| 6 | 11.9 | 12.6 |
| 7 | 11.9 | 11.8 |
| 8 | 10.9 | 11.2 |
| 9 | 10.9 | 10.6 |
| 10 | 10.3 | 10.2 |
| 11 | 9.9 | 9.8 |
Retention curve: observed and power fit. The power curve (a = 0.27, b = 0.43) is close to the observed one over all 11 observed months, R² = 0.99; beyond that it is used only as a forecast. Axis: life month of the cohort.
Data as a table
| Measure | LTV (CM2), RUB |
|---|---|
| 3 | 2,194 |
| 6 | 2,712 |
| 9 | 3,116 |
| 12 | 3,452 |
| 18 | 3,997 |
| 24 | 4,429 |
LTV by horizon (curve forecast). LTV growth slows: from 12 to 24 months it adds RUB 978 against RUB 3,452 accumulated in the first 12. Axis: horizon, months.
- Customer share, %
- Revenue share, %
Data as a table
| Measure | Customer share, % | Revenue share, % |
|---|---|---|
| Champions | 11 | 25 |
| Loyal | 15.7 | 28.5 |
| New | 19.3 | 9.1 |
| At risk | 9.3 | 9 |
| Middle | 23 | 18.5 |
| Lost | 21.7 | 9.9 |
RFM: customer share and revenue share by segment. Champions and loyal customers are 27% of customers and 54% of revenue; the lost are 22% of customers and 10% of revenue.
Specifics by category
- Apparel
- Repeat purchases follow seasons, so monthly retention has spring and autumn peaks; count age in seasons or add a seasonal diagonal.
- Footwear
- A purchase once a season or less: early retention is low, the curve long; the LTV horizon is 24 months or more and the repeat purchase is often in an adjacent category.
- Accessories
- A high share of gift purchases: the recipient is not the customer and repeats fall on holiday dates; cohorts by first-purchase month mix gifts with self-purchases.
- Kids
- The buyer is a parent and the need is tied to child growth and school: child age cohorts are more useful than calendar ones; a segment is lost as the child grows.
- Men
- More even and higher retention on basics; RFM identifies a small set of loyal customers with a large revenue share; promotion dependence is lower.
- Women
- More purchases per customer and higher promotion sensitivity: cohorts acquired through promotions retain worse; slicing cohorts by channel and discount share is mandatory.
Common mistakes
- A simple average over cohorts is used → small cohorts distort the curve → weight by size.
- Cohorts of different channels are mixed → good CRM retention masks poor paid channels → slice by channel.
- The curve is extrapolated far beyond the observed horizon → LTV rests on an assumption → show observed and forecast parts separately.
- Anonymous receipts are ignored → retention understated → estimate the share and state it in the report.
- RFM segments are built without actions and control → mailings bring no effect → an action and a control group for each segment.
Russian specifics
- Storing and processing customer data for cohorts and RFM is allowed with a legal basis and consent, and data of Russian citizens is stored in Russia; for analytics anonymized identifiers are preferable. Rule: Law 152-FZ of 27.07.2006: Federal Law of 27.07.2006 No. 152-FZ on personal data (rule check not completed).
- Mailings by segment require recipient consent and ad labeling; advertising mailings are subject to the advertising law. Rule: Law 38-FZ of 13.03.2006: Federal Law of 13.03.2006 No. 38-FZ on advertising (rule check not completed).
- A loyalty program and bonuses (points) create an obligation to the customer: accounting for and measuring the bonus liability affects revenue and the balance sheet. Rule: Order 56n of 16.05.2025: FSBU 9/2025 Income (Ministry of Finance order of 16.05.2025 No. 56n) (rule check not completed).