Quick answer. There are three ways to predict payback for a subscription app: naive extrapolation (safe but usually too slow), curve-fitting a power-law curve to early revenue data (accurate once enough days of real data exist), and prior-based benchmarks by monetization type and geography (for brand-new campaigns with no data at all). Start every campaign on priors, then graduate to fitting as real data accumulates.

Every UA dashboard has an LTV number on it. Almost none of them tell you how much to trust it.

This is not the beginner version of LTV forecasting. If you want the fundamentals, the funnel from Day 1 signals to a payback window, I already wrote a separate piece on how to forecast LTV and payback. This one is narrower and more specific to subscriptions: the three actual ways to build the prediction, the real math behind each, and why picking the wrong one can put your payback date more than a month off.

What a payback model actually predicts

A payback model answers one question: on which day will this cohort's cumulative subscription revenue equal what you spent to acquire it? Get that day right and you know exactly when a campaign turns from a cost into a profit center. Get it wrong and you either scale a campaign that is quietly losing money, or kill one that was about to pay off.

The number itself is easy to produce, any spreadsheet can divide revenue by spend. The part almost nobody checks is how much to trust that number on day 3 versus day 30. That is what separates a real payback model from a guess with a formula attached to it.

The three types, and the real math behind each

1. Naive extrapolation. Take whatever revenue a cohort has made so far and assume the daily rate keeps decaying at a fixed pace, rather than trying to model how it actually grows. In practice: sum the observed revenue, then extend the tail with a fixed decay rate applied day over day. It takes one line of code and zero data science. The tradeoff is that the decay rate is a guess, not something fit to your actual cohort, so it is almost always too conservative. It is the safe model, and the slow one.

2. Curve-fitting. Instead of guessing a decay rate, fit an actual curve to the revenue you have observed, usually a power law or a logarithmic curve, and keep whichever one fits your real numbers with the least error. A production LTV engine I run does exactly this: it tries a few candidate curve shapes, keeps the lowest-error one, and will refit with more weight on recent days if a cohort's behavior is shifting mid-flight. It is genuinely accurate once you have enough real data points spread over enough days. With two data points from the last two days, it will confidently draw you a beautiful curve that means nothing.

3. Prior-based benchmarks. For a campaign that launched yesterday, there is no curve to fit yet, so you borrow one. A shape based on the monetization type (subscriptions renew in a predictable ladder, in-app purchases often skew whale-shaped, ad-monetized apps front-load a spike and decay), the geography (a US cohort typically monetizes at a different level and renewal rate than a tier-3 country), and the traffic source (organic users usually convert and retain better than paid ones). None of it is specific to your campaign. All of it beats having nothing on day one.

Same 7 days of data, two different answers

$0$4$8$12Cumulative revenue per installDay 0Day 30Day 60Day 90Spend to recover: $8/installDay 45Day 80Fitted curveNaive decay
Same 7 days of observed revenue, two different projections. The fitted power-law curve says this cohort breaks even on day 45. The naive decay model, run on the exact same data, says day 80: a 35-day gap that decides whether you scale, pause, or kill the campaign this week or a month from now.
Fitted curve (power-law regression)Naive projection (flat decay)Observed, days 1, 3, 7

The chart uses one small, real-shaped example: a cohort that costs $8 per install, with revenue observed through day 7. Both models agree on what already happened. They disagree hard on what happens next, and that disagreement is not a rounding error, it is over a month of deciding whether to scale, pause, or kill a campaign based on which model you happened to trust.

Which one you should actually use

Use the type that matches how much data you actually have, not the one that is easiest to build. Start every new campaign on prior-based benchmarks, because on day one that is genuinely the best information available. As real revenue comes in, switch to curve-fitting, and trust it more as the data spans more days and more cohorts, not just more raw revenue.

The part worth stealing even if you never build any of this yourself: attach a confidence level to every payback number, not just the number itself. A payback estimate built on two data points and one built on thirty should never look the same on your dashboard, even when they happen to land on the same day.

I built exactly this, the switch from priors to curve-fitting and the confidence scoring, inside uacopilot.ai, a UA analytics product I run. This is not theory I read somewhere, it is the model deciding real budget calls on real campaigns.

Key takeaways

  • A payback model is only as trustworthy as the data behind it, so its confidence level matters as much as its number.
  • Naive decay models are safe but usually late; curve-fits are accurate but need real data spanning real days, not just two points.
  • Start new campaigns on prior-based benchmarks by monetization type and geo, then graduate to a fitted curve as revenue data accumulates.

Frequently asked questions

How many data points do I need before I trust a curve fit?
As a rule of thumb, at least four observed points spanning at least a week. Fewer than that and a curve fit is really a prior-based guess wearing a fitted formula. Treat anything fit on two or three early days as low confidence, no matter how clean the curve looks.
Why do naive projections usually run late instead of early?
Because a fixed decay rate assumes the cohort is already slowing down as fast as it ever will. Real subscription cohorts often keep converting trial users and renewing subscribers well past the point a flat decay curve expects, so the naive model tends to underestimate future revenue and push the payback date out.
Does more days of data always mean a more accurate model?
More days of real data make curve-fitting more accurate, but a projection reaching further out than the data actually covers is extrapolating into the unknown, not gaining certainty. A common practical rule: do not trust a projection more than about six times further out than the days you have actually observed.