Customer Lifetime Value (LTV)

Customer lifetime value (LTV, also written CLV or CLTV) is the expected net value a customer will generate for a business over the rest of their relationship with it, usually expressed as the discounted sum of future margin. It is a forecast, not an accounting figure. LTV sets the ceiling on what it is rational to pay to acquire a customer and guides how much to invest in retaining one, which makes it a central number in subscription businesses, e-commerce and free-to-play games. It is also easy to get wrong: spending is highly skewed, the horizon is long, and the customers observed so far are not a fair sample of those to come.

Definitions

Four choices define any LTV figure, and two numbers that differ on them are not comparable.

• Revenue or margin. Gross revenue overstates value; contribution margin after fees, refunds and cost of service is the meaningful base.
• Horizon. "Lifetime" is in practice truncated to a fixed window such as 90, 180 or 365 days; an infinite horizon requires a discount rate.
• Historical or predictive. Realised value to date is a fact; predicted LTV (pLTV) is a model output.
• Individual or cohort. Cohort averages are more stable than per-customer estimates and suffice for many decisions.

The textbook subscription shortcut, average margin per period divided by the churn rate, assumes that churn is constant over time and identical across customers. Neither usually holds: early churn is typically much higher than late churn.

Classic Probabilistic Models

The foundational models address the non-contractual case, where customers buy when they like and never announce that they have left. They need only three numbers per customer: recency, frequency and the length of time observed.

Model	Source	Behavioural story
Pareto/NBD	Schmittlein, Morrison and Colombo, Management Science, 1987	While "alive," a customer buys at a steady random (Poisson) rate that varies across customers; each customer drops out after an unobserved lifetime, also heterogeneous
BG/NBD	Fader, Hardie and Lee, Marketing Science, 2005	Same purchasing process, but dropout can occur only immediately after a purchase, with a probability that varies across customers
Gamma-Gamma	Fader, Hardie and Lee, "RFM and CLV," 2005 (adapting Colombo and Jiang, 1999)	Each customer has a latent average transaction value; observed spend varies around it

Fader, Hardie and Lee introduced BG/NBD as "a slight variation in the behavioral story" of Pareto/NBD that is far easier to estimate, noting that its parameters can be obtained in a spreadsheet and that the two models give very similar results across a wide range of purchasing environments. Either yields, per customer, the probability of still being active and the expected number of future transactions.

The Gamma-Gamma model supplies the monetary side. It assumes that the value of a transaction is independent of how often a customer buys, so expected transactions and expected spend per transaction can be multiplied. The authors found a correlation of only 0.11 between average transaction value and number of transactions among repeat buyers at an online music retailer; the assumption should be re-checked on any new dataset. Contractual settings use survival and retention-curve models instead; there, as the PyMC-Marketing documentation puts it, "the company always knows the status of the customer relationship."

These models remain useful: they work with little data and extrapolate beyond the observed window in a principled way. In their basic forms they ignore covariates, assume stationary behaviour, and say little about a customer who has only just arrived.

Machine-Learning Approaches

The alternative is to treat LTV as supervised regression: compute features from a customer's first days and train a model, typically gradient-boosted trees or a neural network, to predict value over the following months. This uses rich behavioural signals and handles early prediction directly, but needs a full outcome window of history before any label exists.

The statistical obstacle is the shape of the target: many customers are worth nothing further, and a few are worth a great deal. Wang, Liu and Miao (2019) proposed modelling LTV as a zero-inflated lognormal distribution, a point mass at zero mixed with a lognormal, so that one network jointly learns the probability of any future spend and its size. The paper recommends a normalised Gini coefficient for ranking and decile charts for calibration.

Sequence models and behavioral foundation models push further by learning from raw event logs. Chen et al. (2018) reported that convolutional networks on raw player data outperformed Pareto/NBD-style models in free-to-play games, where, they noted, up to half of revenue can come from around 2% of players. That is a single study. Tabular foundation models are a newer option, with little published LTV-specific evidence as of October 2026.

Uses

Acquisition bidding. Predicted LTV by channel and campaign, compared with acquisition cost, determines budgets and payback periods. Where advertising platforms bid on predicted value, forecasts are needed within days of install or first purchase, when very little behaviour has been observed.

Live-service products. In live-service games and subscription apps, LTV estimates segment customers, set the budget for retention offers, and serve as the long-run metric against which changes to pricing, content and the in-game economy are judged.

Pitfalls

• Extrapolating from young cohorts. A 365-day forecast built on 30 days of data is mostly assumption; back-test against cohorts old enough to have real outcomes.
• Survivorship and mix shift. Early and organically acquired customers are usually more valuable than those reached later through paid scaling.
• Self-fulfilling predictions. If high-predicted-value customers receive better treatment, their higher realised value partly reflects the treatment; randomised holdouts separate the two.
• Prediction is not causation. Knowing a customer's expected value does not say what an intervention would add to it. That is an uplift question, a limitation shared with churn prediction.