LTV prediction for subscription apps: how Prediction Explorer works

The realized part is the same as Cohort Explorer's realized LTV per customer: cumulative revenue of the cohort from each customer's start through the end of each month of their own age, divided by the cohort size, net of refunds and in US dollars at the purchase-date rate. Month 0 is the first month. The cohorting_date selector groups customers by new_customers (the default), initial_conversions or new_paying_customers.

For each month n, RevenueDot adds up the month n values of all cohorts that have completed both month n−1 and month n, and divides by the sum of their month n−1 values. That ratio is the growth factor for month n. A cohort's missing months are its last realized value multiplied by the growth factors in order. If no cohort has completed a month yet, its factor is 1, so the projection stays flat. Predicted cells carry predicted: true and incomplete: true in the API.

This is a projection from your own older cohorts. It does not use data from other apps, a survival model, seasonality or knowledge of a price change. It assumes newer cohorts grow like older ones. Sandbox purchases, granted access and Family Sharing purchases are excluded, and ad revenue is not included. Filter by app, store, product, product duration, offering, country, platform and app version, but not segment.

Find it in the dashboard under Analytics > Charts, or call GET /v2/projects/{project_id}/charts/prediction_explorer. The response has the same shape as RevenueCat's Charts API: a cohort table with periods[0] as the cohort size.

Illustrative numbers with two cohorts. Cohort A has completed month 1. Cohort B has not.

Growth for month 1 = 3.00 ÷ 2.00 = 1.5, so cohort B's predicted month 1 = 2.50 × 1.5 = $3.75. With more cohorts, the month 1 values are added up before the division.

Waiting a year to learn what a customer is worth is too slow when you buy traffic every week. A projection from the cohorts you already have lets you estimate whether a channel will pay back by month 6 or month 12, and keep or cut spend sooner.

Treat it as an estimate. It works best when your older cohorts look alike and your product has not changed much. Compare predicted cells of a recent cohort with the realized columns as they arrive, and learn how far off the projection is for your own app.

How to read it

• Predicted values are flagged predicted: true in the API. Keep them apart from realized values in any report.
• A new price, paywall or onboarding breaks the assumption that new cohorts grow like old ones.
• With only a few months of history, later months are held flat because no older cohort has reached them.
• Small cohorts are noisy. A few big purchases move the growth factor.
• Per-customer values include people who never pay. Use the new paying customers cohort to leave them out.