Cohort-Based LTV

Cohort-based LTV estimates lifetime value by grouping customers by when they joined and watching what each group actually does, month by month. Instead of assuming a constant churn rate, it observes the cohort.

Define a cohort by a common starting event, such as first paid activation in January. For each period since that event, total its revenue or contribution and divide by the original cohort size. Accumulate those amounts over a stated horizon. Keep observed periods separate from any forecast of future periods. A temporary flat section of the curve does not establish that all future value is zero: seasonal purchases, delayed renewals or reactivation can change it.

\[ \text{LTV} = \sum_{t=0}^{T} \frac{\text{average revenue per customer in month } t}{(1 + r)^t} \]

Here t = 0 is the acquisition period and T is the last included period, so the sum contains T + 1 periods. The discount rate r must match that period length. This version measures revenue LTV; use contribution after a defined cost boundary when that is the objective. If T is simply the last observed period, the result is accumulated value through that age, not automatically full lifetime value. The LTV entry explains timing and discounting.

Keep the original customer denominator

Suppose a cohort starts with 100 customers. In month one, 74 remain active and generate 2,960 of revenue. Revenue per active customer is 40, while revenue per originally acquired customer is 29.60. In a later month, 43 active customers generating 2,107 produce 49 per active customer but only 21.07 per original customer. The first metric describes spending among survivors. The second preserves departures in the cohort value calculation.

Summing per-active-customer averages would combine different survivor populations across months. It does not measure lifetime revenue per acquired customer. Keep the original denominator fixed, include zero-revenue periods for inactive members, and handle refunds or returning customers consistently. Neither a rising survivor average nor a declining cohort total identifies its cause without examining individual behavior and composition.

Label the time axis before adding values

For a separate illustrative cohort, assume revenue per original customer is 100 at month zero, followed by 40, 38, 35, 32, 30 and 28. The total through month three is 100 + 40 + 38 + 35 = 213. Six periods, months zero through five, sum to 275. Through month six there are seven periods, summing to 303. Label “month since acquisition” and the included endpoints so a six-month figure cannot silently include seven periods.

Compare cohorts at the same age. A January cohort observed for twelve months and an October cohort observed for three months cannot be compared by their current cumulative totals as if they had equal follow-up. In a cohort table, unobserved future cells are missing, not zeros. Show the three-month values for a fair observed-age comparison, or clearly label an independently validated forecast to a common later horizon.

The unobserved tail needs a model

A declining stream can have very different totals depending on what happens after observation ends. Starting after the observed month-six value of 28, suppose future monthly revenue per original customer falls by 10% each month. The next amount is 25.20 and the undiscounted infinite future sum is 25.20/0.10 = 252. If the decline is only 5%, the next amount is 26.60 and the future sum is 26.60/0.05 = 532. These are two hypothetical tails, not fits to the observed data. They make the same already observed total of 303 lead to projected totals of 555 or 835.

A fitted exponential or power-law curve also embeds assumptions about how the tail behaves. A small difference in fit over a short observed interval need not support a large difference in long-run value. Use mature historical cohorts to test forecasts made using only their early periods, and consider the product’s renewal cycle, changing costs and plausible retention mechanisms. A power-law tail can even have a divergent infinite sum under some parameters, so a finite horizon and a clearly stated discount convention can be essential.

Report the observed amount, forecast amount and horizon separately, together with sensitivity to plausible alternatives. If the tail dominates the estimate, the decision is particularly dependent on modeling assumptions. If no credible tail model is available, a fixed-horizon observed value is more informative than an unsupported lifetime total.

What it is better than, and where it stops

Compared with a constant-churn shortcut, cohort tracking shows how revenue and retention evolve with tenure. The customer lifetime entry explains why mixed customer populations need care. Cohort comparisons help separate acquisition timing and tenure from aggregate mix, but they do not by themselves establish that a product change caused better outcomes. Channel, seasonality and customer characteristics can differ between cohorts.

Its limits follow from the same structure. A cohort has to age before it says anything, so the most recent cohorts — the ones a decision is usually about — are the least observed. Cohorts differ for reasons other than product changes, including channel mix and seasonality, so a worse cohort is not automatically a worse product. And the projection has to be stated as one: the method’s credibility comes from the observed months, and it is easy to spend that credibility on the fitted ones.

Where decisions concern recent customers, predictive LTV can use early signals to forecast a specified horizon. That approach still needs historical evidence and validation; it complements cohort tracking rather than removing the uncertainty of unobserved outcomes.

References

• Stripe Atlas, SaaS business model — revenue, cohorts and retention