A cohort-based survival model groups token holders by the month they arrived and tracks what share of each group is still holding at every month after entry, instead of dividing one churn rate across the whole holder base. The machinery is borrowed from survival analysis, with a wallet standing in for a subject and a defined exit event standing in for the endpoint. In a token model it answers one question a blended retention number cannot: is the cohort your last incentive campaign bought behaving worse than the one before it.
An aggregate holder count can rise while every recent cohort decays faster than the last, because new arrivals mask the decay of the old ones. Cohorts separate those two movements, and the second one is the one that decides whether an incentive budget is buying anything durable.
One curve per entry cohort, not one churn rate
Split holders by acquisition month. The June airdrop cohort, the July listing cohort, the August liquidity mining cohort. For each, count how many are still holding at month one, month two, month three, and plot that as a curve falling from 100 percent. You now have a set of curves that can be laid over each other from a common origin, which is the entire point: cohort three can be compared to cohort one at the same age rather than at the same date.
The estimator underneath is Kaplan-Meier, the product-limit method that computes survival as a running product of period-by-period conditional survival probabilities.1 Nothing about it is crypto-specific. It is the same arithmetic used in clinical follow-up studies, which is why the definitions carry over cleanly and why the failure modes do too.
Censoring, and why it is not a technicality
Most of your wallets have not exited yet. A wallet that joined two months ago and is still holding tells you it survived two months and nothing at all about month twelve. Survival analysis handles this as right censoring: the observation is incomplete rather than missing, and the wallet stays in the risk set for the months you observed it and drops out afterwards.2
Get this wrong in either direction and the model lies to you. Treat unexited wallets as survivors and you overstate retention badly, because recent cohorts have barely had time to leave. Drop them and you throw away most of your data and bias the result toward whoever exited early. The Kaplan-Meier estimator exists to handle exactly this case, and it is the reason to reach for survival analysis rather than a spreadsheet retention grid.3
Defining the exit event is the hard part, and wallets are not people
In a clinical study the event is unambiguous. In a token economy you have to define it, and the definition changes the answer. Balance falls to zero. Balance falls below the amount they were airdropped. No transaction for ninety days. Unstaked and moved to an exchange deposit address. Each is defensible and each produces a different curve, so the definition belongs in the model documentation, not in a footnote.
The unit is the bigger problem. A wallet is not a person. One holder can run ten addresses, an airdrop farmer runs thousands, and a custodial exchange holds tens of thousands of real holders behind a single address that never moves. Across the token models we build, the fix is to weight the curve by value rather than count it by address, so that survival measures the share of a cohort's balance still held rather than the share of its addresses still non-empty. That reweighting usually changes the story.
What the curves change at design time
Three decisions, directly. Incentive budget, because a cohort whose value-weighted survival collapses in the first two months is telling you the campaign bought sell pressure with extra steps, and no further spend on that channel is justified until the mechanism changes. Emission pacing, because the rate at which the holder base decays sets how much new demand emissions have to attract just to stand still. And lock design, because a cliff that lands where a cohort's curve is already falling stacks two exits on the same month.
The pattern worth watching is not the level of any single curve but the drift between successive ones. Rising curves across cohorts means the product is doing the retaining. Falling curves with steady headline holder counts means the incentives are, and that is a treasury expense with a fixed end date rather than a business.
Common questions
What is a cohort-based survival model in tokenomics?
It groups token holders by the month they arrived and tracks the share of each group still holding at each month of age, using the Kaplan-Meier estimator from survival analysis.1 Because cohorts are compared at the same age rather than the same calendar date, it separates genuine retention improvement from the masking effect of new arrivals, which a blended churn rate cannot do.
Why use survival analysis instead of a simple retention rate?
Because most of your holders have not exited yet, and a simple rate has to either ignore them or count them as survivors. Survival analysis treats them as right-censored: they contribute to the months you observed and leave the risk set afterwards.2 That handling is what makes retention numbers comparable across cohorts of different ages, which is the whole reason to run the analysis.
What counts as churn for a token holder?
You have to define it, and the definition changes the curve. Common choices are balance reaching zero, balance falling below the amount received, a period of inactivity, or a transfer to an exchange deposit address. Whichever you pick, state it alongside the results and keep it constant across cohorts. Weighting by value rather than by address count also matters, since one person can hold many wallets and one custodial address can hold many people.
See Tokenomics Design for how this applies in practice.
Sources
- STAT 331 Unit 3: The Kaplan-Meier estimator (course notes)
Stanford University, Department of Statistics
Formal definition of the product-limit estimator, risk sets and event times. - The Kaplan-Meier Method for Estimating and Comparing Survival
PubMed Central, US National Institutes of Health
Peer-reviewed methodological tutorial covering right censoring and the interpretation of survival curves. - Survival Analysis Using S/R (course script)
ETH Zurich, Seminar for Statistics
Treatment of censoring mechanisms and estimator construction, used here for the censoring handling that a plain retention grid cannot represent.
Last reviewed 2026-08
Know the terms but not sure how they apply to your project? That is what an engagement is for. We design, document, and stress-test the whole token economy inside the Tokenomics Data Room.
100+ projects advised. Complete tokenomics in 4 to 6 weeks.