LTV projection: predicting from your history, not a fitted curve

By the HonestTag team · Published July 12, 2026

Young cohorts pose an awkward question: you need to know what a March customer will be worth by month 6, and March was two months ago. One answer is a statistical model. HonestTag answers with your own history.

How the projection works

The method is cohort-matching. To project a young cohort one month forward, HonestTag looks at every older cohort of 30 or more customers that actually reached that month, computes each one's month-over-month growth ratio of cumulative revenue per customer, and takes the median of those ratios. That median chains onto the young cohort's last complete month. Repeat for the next month, and the next, each step drawing its median from the older cohorts that really lived through it.

That is the whole formula, and it is published in numbers on the in-app methodology page. Every input is a cohort you can see on the same screen. If you doubt a projected figure, you can trace it: these donors, these ratios, this median. Projections ship on the Truth+ tier and up. Correctness is never paywalled; breadth and judgment layers are.

The floors: no donors, no number

Each projected step needs at least 3 donor cohorts of 30+ customers covering that month. Below the floor, the cell shows unlock copy ("projections unlock when 3 cohorts of 30+ customers reach month N") and never a number. The chain stops at the first unfloored step; it does not skip over a thin month and keep going on the far side.

This is fail-loud design. A projection built on one donor is just that donor's history wearing a new date. Rather than quietly degrade, the report tells you exactly what it is missing and fills in on its own as your cohorts age.

The horizon: never past your own history

Projections never extend beyond the deepest qualifying donor's observed depth, and never past month 12, period. If your oldest 30-customer cohort has reached month 8, projections stop at month 8. The store is never shown a future it has not already exhibited somewhere in its own history.

This rule costs us impressive-looking charts. A store with six months of data will not see a 24-month LTV figure here, because no honest one exists. What it sees instead is a horizon that extends every month, on its own, as real history accumulates.

You can always tell a projection from a fact

Every projected cell is visibly distinct: a tilde-prefixed figure, a low-to-high range beside it, and a confidence badge. Actuals and projections are never summed into one number. The summary is always two figures, "observed to month L" and "projected months L+1 onward", side by side, never a single total. And when a payback crossing lands in the projected region rather than the observed one, the payback column says "projected" explicitly, because spending against a projected payback as if it were observed is exactly the mistake the label exists to prevent.

Confidence badges from donor dispersion

How much your older cohorts agree determines how the projection renders, on published thresholds. When donors are few or their paths vary, the projection carries a "directional" badge: useful for planning, not for precision. When cohorts disagree strongly, the badge reads "volatile": read the projection as its range, not its middle figure, because a single number would claim a consensus your history does not contain.

Channels project only from themselves

Where cohort curves split by acquisition channel, projections respect the split: a channel's cells project only from that same channel's own qualifying donor cells. A Google slice is never projected from blended history, because blended history describes a customer mix the Google slice does not have. And the "(before channel tracking)" bucket is never projected at all. Its early months predate tracking, so projecting off it would be a fake number twice over.

What the dishonest version looks like

Most attribution dashboards that show "predicted LTV" share a few habits worth checking for. Fitted exponential decay curves with parameters nobody can audit, where the chart looks smooth because the formula guarantees smoothness, not because your customers behave smoothly. Predicted LTV blended silently into reported LTV, so you can no longer tell which part of the number happened. Projections that extrapolate to month 36 for a store with five months of orders, past anything the store ever did. And machine-learned forecasts presented without their inputs, where "trust us" is the methodology page.

What a projection is (and isn't) for

1. It is a forward sanity-check, not a settled number. When the projected curve says a cohort clears its NCAC by month 5 and a ceiling assumed month 3, that gap is visible now rather than in the spring. The projection surfaces it; it does not resolve it.
2. A "volatile" badge is a range, not a decoration. It means the cohorts disagree about the future; the band it shows is the spread, which is wider than any single midpoint.
3. A projected payback and an observed one are labeled differently on purpose. Verdicts and observed months are the record of what has happened; a projection describes a future the store has not yet reached.

Frequently asked questions