How to Predict Customer Lifetime Value Before You Have Years of Cohort Data

Paid acquisition runs on one number: how much a new customer is worth. A historical LTV only becomes reliable once a cohort has aged for years, and no growth team can pause its bidding that long. Without a prediction, CAC targets rest on a guess that only gets checked years later.

You can predict customer lifetime value early by fitting a probability model to the behavior you already have and projecting it forward. Subscription businesses fit a retention curve to their observed renewal periods. Repeat-purchase businesses model recency and frequency. Customers with one order need a model trained on first-purchase signals. Test every model against a held-out period.

What is predicted customer lifetime value, and how does it differ from historical LTV?

Historical LTV adds up what a cohort has already spent. It is a fact, but it is only complete once the cohort has stopped buying, which can take years. Predicted LTV is an estimate of the value a customer or cohort will generate in the future, based on a model of how customers like them behave.

The prediction is only as useful as the value it measures. Build it on contribution margin, after product cost, shipping, payment fees and returns, so it can be compared with CAC directly. Our guide to ecommerce LTV:CAC covers how to build both sides of that ratio.

Why can't you just extrapolate the first few months of a cohort?

Because early retention data does not follow a straight line, and simple curve fits break when you push them past the data. Peter Fader and Bruce Hardie tested this in the Journal of Interactive Marketing in 2007. They fit linear, quadratic and exponential curves to cohort survival data and projected them forward. In their words, all three models "break down dramatically." The linear model underestimated survival in year 12 by 81%, and the quadratic curve eventually predicted that the number of surviving customers would rise, which cannot happen.

The reason is heterogeneity. Customers differ in how likely they are to churn, and the high-churn customers leave first. Fader and Hardie write that rising retention rates over time are "simply due to heterogeneity," and that this is "not a story of individual customers becoming increasingly loyal." The practical consequence: if you take the churn rate from a cohort's first months and hold it constant, your projection will show too few long-tenured customers, because the customers who remain churn less than the cohort average did at the start.

Which LTV prediction method fits your business model?

The right method depends on two things: whether customers hold a contract or subscription, so you can see the moment they leave, and whether you have any purchase history for the customer yet.

1. Subscription businesses: the shifted-beta-geometric model

In a contractual setting, such as a monthly or annual subscription, you observe churn directly. The shifted-beta-geometric (sBG) model from the 2007 Fader and Hardie paper assumes each customer has a constant probability of cancelling at each renewal, and that this probability varies across customers following a beta distribution. That second assumption is what lets the model reproduce rising retention rates.

The input is simple: the share of each cohort still active at each renewal period. The open-source PyMC-Marketing library includes an sBG implementation, and its tutorial fits the model on the first 7 of 13 periods and checks the projection against the rest.

Limits: the model assumes a customer's churn probability does not change over time, so a price increase, a product change or a new cancellation flow can make an older fit wrong. It also works in discrete renewal periods, so it suits billing cycles better than continuous usage.

2. Repeat-purchase businesses: BG/NBD and Gamma-Gamma

An ecommerce business without subscriptions never sees a customer cancel. A customer who has not ordered in 90 days may have left or may simply be between orders. The BG/NBD model, published by Fader, Hardie and Ka Lok Lee in Marketing Science in 2005, handles this non-contractual setting. It needs three numbers per customer: frequency (the number of repeat purchases), recency (the time of the most recent purchase), and T (the time since the first purchase). From those it predicts expected future purchases and the probability that each customer is still active.

The model does not need years of data. In the original paper, the authors calibrated it on 39 weeks of purchases from 2,357 customers who first bought from the online retailer CDNOW in the first quarter of 1997, then forecast the following 39 weeks. The aggregate forecast came in 4% under the actual number of repeat transactions.

Purchase counts are half the answer. The Gamma-Gamma model estimates average spend per transaction, and a 2005 paper by the same authors pairs a purchase model with a gamma-gamma sub-model for dollars per transaction to link recency, frequency and monetary value to CLV. Its key assumption is that there is no relationship between how much a customer spends per order and how often they buy. The PyMC-Marketing tutorial checks this by measuring the correlation between average spend and purchase frequency, and it fits the model on repeat customers only.

Limits: both models assume past behavior predicts future behavior. They do not know about your next promotion, a change in channel mix or a shift in the type of customer you acquire. Refit them when any of those change.

3. New customers: first-purchase signals and machine learning

Recency and frequency models have a blind spot at the moment you need a prediction most. Xiaojing Wang, Tianqi Liu and Jingang Miao of Google state it directly in a 2019 paper: the BTYD model family, which includes BG/NBD, "does not apply to new customers because it uses frequency and recency to differentiate customers." Every new customer has the same frequency and recency.

Their approach predicts value from customer attributes and purchase behavior instead. They model LTV as a mix of a point mass at zero and a lognormal distribution, which they call zero-inflated lognormal (ZILN). That handles two facts that break a standard regression: a significant fraction of customers buy only once, and a small set of high-value customers spend orders of magnitude more than a typical one. In one of their experiments, the model predicts each customer's total purchase value in the 12 months following the initial purchase. Google publishes the code as an open-source package.

Features available at the first order, such as first basket value, product category, acquisition channel and whether a discount was used, are the raw material for this kind of model. Which ones carry signal is specific to each business, so test them rather than assume them.

4. Platform predictions: Google Analytics 4 predictive metrics

GA4 calculates predictive metrics for eligible properties. Predicted revenue is "the revenue expected from all purchase key events within the next 28 days from a user who was active in the last 28 days." Purchase probability and churn probability cover 7-day windows.

These are useful for audiences and short-term targeting, but they cover 28 days, which is far short of lifetime. Treat them as an early signal that feeds your own model.

How much data do you need before an LTV prediction is usable?

There is no universal minimum, but the methods above set concrete floors:

• GA4 predictive metrics: at least 1,000 returning users who triggered the relevant condition and 1,000 who did not, over a seven-day period within the last 28 days. Purchase events must include the value and currency parameters.
• sBG for subscriptions: the beta distribution has two parameters, so the model needs at least two observed renewal periods to fit at all. More periods make the projection steadier.
• BG/NBD: the original paper worked from 39 weeks of data on 2,357 customers. Your own holdout test decides whether that is enough for your data.
• Gamma-Gamma: customers with at least one repeat purchase, since the tutorial fit excludes one-time buyers.

The real test is whether your model can forecast a period it has not seen. If it cannot, try a different model, or split out cohorts that behave differently, before you wait for more history.

How do you check that a predicted LTV is accurate?

Use the same design the research uses: calibrate on an earlier window, forecast a later one you held back, and compare. The Wang, Liu and Miao paper recommends two checks for machine-learning models. The normalized Gini coefficient measures whether the model ranks customers correctly. Decile charts compare predicted and actual value for each tenth of customers, which shows whether the predicted dollar amounts are right.

An illustration, with made-up numbers: you fit a BG/NBD model on a cohort's first 26 weeks and hold out the next 13. The model forecasts 1,150 repeat orders for the holdout window, and the cohort places 1,060. That is an overforecast of about 8%. Before you use the model to set bids, check whether the miss is spread evenly or concentrated in one acquisition channel. A bias in one channel will push budget in the wrong direction even when the total looks close.

Refit on a schedule, and refit immediately after a pricing change, a new offer or a shift in channel mix. A prediction calibrated on last year's customers describes last year's customers.

How do you use predicted LTV to set acquisition budgets?

Start with the prediction's range. Say a model predicts a new cohort's 24-month contribution margin at $120 per customer, with a plausible range of $90 to $150 (an illustration). Setting the CAC ceiling from the $90 end means that if the model proves optimistic, the cohort still breaks even on contribution margin. As cohorts age and the range narrows, raise the ceiling. For timing, compare the prediction with your CAC payback period, since cash recovery matters as much as the lifetime total.

Watch for drift as you scale. New spend tends to reach customers who differ from the ones the model learned from. We cover how that plays out in why LTV:CAC deteriorates as subscription businesses scale. The stakes grow with volume: Exactius has acquired more than 12 million members and customers, and at that scale a small bias in predicted LTV repeats across every bid.

Predictions can also feed ad platforms directly. Google Ads customer lifecycle goals include "New customers (high value)," which you define with a customer list segment and assign a value higher than the standard new customer value. Google notes that what counts as high value is business specific, and a predicted LTV score is one way to build that list. On July 22, 2026, Google announced version 25 of the Google Ads API, which added API support for a retention goal that Google says is "designed to drive customer loyalty and high lifetime value (LTV)" by optimizing for loyalty program members. Some re-engagement goals, such as lapsed customers, are still labeled beta.

What are the most common mistakes in early LTV prediction?

• Predicting revenue instead of margin. A high-revenue customer who buys only discounted, low-margin products can be worth less than a smaller spender.
• Using one model for every cohort. Customers from different channels, offers or price points can behave very differently. Fit them separately, or include those differences as features.
• Holding early churn constant. As Fader and Hardie show, retention rates rise as high-churn customers leave, so a constant early churn rate understates long-run survival.
• Treating a 28-day platform prediction as lifetime value. GA4 predicted revenue covers 28 days by definition.
• Skipping the holdout test. A model that fits past data well can still forecast poorly. Only a held-out period shows which kind you have.

Sources

• Fader and Hardie, How to Project Customer Retention, Journal of Interactive Marketing (2007)
• Fader, Hardie, and Lee, "Counting Your Customers" the Easy Way: An Alternative to the Pareto/NBD Model, Marketing Science (2005)
• Fader, Hardie, and Lee, RFM and CLV: Using Iso-Value Curves for Customer Base Analysis (abstract)
• Wang, Liu, and Miao, A Deep Probabilistic Model for Customer Lifetime Value Prediction (arXiv, 2019)
• Google, lifetime_value open-source package (GitHub)
• PyMC-Marketing documentation, BG/NBD model
• PyMC-Marketing documentation, Gamma-Gamma model
• PyMC-Marketing documentation, shifted beta geometric model
• Google Analytics Help, predictive metrics
• Google Ads Help, about customer lifecycle goals
• Google Ads Help, activate customer lifecycle goals
• Google Ads Developer Blog, announcing v25 of the Google Ads API (July 22, 2026)

If you want to know how far your current cohort data can take an LTV prediction before you set next quarter's CAC targets, book a call and an Exactius operator will review it with you.

Exactius is a full-funnel growth agency accountable for its clients' P&L. Its AI-enabled senior operators provide performance marketing, strategy, creative, and whole-business analytics and data science, engaged one function at a time or as a full team. It serves consumer and B2B companies where paid marketing is a main growth lever, through two practices: one for companies from $5M to $100M and one for companies from $100M to $1B.

FAQ

Frequently asked