Predictive LTV (pLTV)

Predictive LTV, or pLTV, estimates a customer’s future revenue or contribution using information available at a specified prediction time. It addresses the unobserved part of lifetime value. Define whether the target is remaining value, total value including what has already been observed, or value through a fixed age such as twelve months. Those are different targets, even when all are called pLTV.

\[ \text{pLTV}_i = \sum_{t=1}^{T} \mathbb{E}[\text{revenue}_{i,t} \mid X_i] \]

Here i identifies a customer, T is the forecast horizon, and X_i contains information known at the prediction cutoff. The displayed formula is undiscounted future revenue. For contribution, subtract the defined future costs; for present value, discount each period consistently. Forecasts can inform budgets or prioritization, but customer value alone is not the value caused by an advertisement or an intervention.

Several methods serve different data settings. An early-behavior rule estimates future average value for customers with similar observed histories. Cohort-based models use historical age profiles and project their tails. In noncontractual repeat purchasing, Pareto/NBD or BG/NBD models represent purchase activity and unobserved departure; a monetary-value model supplies expected spending per purchase. Supervised regression can instead learn a fixed-horizon target from customer features. Choose based on the relationship type, available history and target, rather than treating one model family as universally appropriate.

Fix the prediction date before building features

For a day-seven forecast of contribution in days eight through 365, feature values must be available by day seven. Future refunds, eventual churn status, and a customer’s full-year purchase count cannot be features. Training examples need the same cutoff and target construction. A customer observed for only two months does not have a complete 365-day target; the missing tail must not be labeled zero.

For a fully supervised fixed-horizon model, use appropriately matured labels. Survival or generative models can use incomplete histories under explicit assumptions, but extrapolation is still a forecast. Separate customers and calendar periods in evaluation as required by deployment. Multiple snapshots of one customer can leak information across a random split, and a model for new acquisition cohorts should be tested on later cohorts rather than only random historical rows.

Distinguish total accuracy, ranking and individual uncertainty

Suppose two customers eventually produce 50 and 150. Predicting 100 for each gets the total of 200 right while failing to distinguish them. Predicting 60 and 140 also gets the total right and captures their order. A correct aggregate therefore does not establish useful individual targeting. Conversely, an aggregate forecast may be the intended product, in which case the constant baseline has value as a comparison rather than being inherently useless.

A conditional expected value is an average over possible future outcomes for similar information. A prediction of 100 does not promise that this individual will return exactly 100. Even a well-specified forecast can have substantial individual error when future behavior is uncertain. Compare absolute or squared errors with a baseline, inspect totals, and evaluate ranking or top-group value only when the decision uses them. A low R² need not imply useless ranking, but the size of any useful lift must be measured on the actual application.

Check calibration by groups of predicted value: among customers predicted around 200, does the average realized target approach 200 as enough outcomes mature? Report group sizes and uncertainty. A model may get the overall average right by overpredicting one group and underpredicting another. Deciles are one diagnostic view, not proof of calibration at every value or within every customer segment.

Choose assumptions that match the purchasing process

BG/NBD and related models are designed for noncontractual transactions where departure is not directly observed. They use a specified purchase process and a dropout process, with variation between customers. Gamma-Gamma monetary models introduce assumptions about positive transaction values and their relation to purchasing frequency. A model designed around repeat purchases should not be transferred unchanged to fixed subscription cancellations, negative net transaction amounts or strong seasonality without checking those assumptions.

For sparse early histories, pooling information across customers can stabilize estimates. A hierarchical model or regularized regression can reduce an extreme estimate based on one purchase toward a supported group mean. But shrinkage is not a mandatory second correction: an already calibrated conditional mean can be made worse by additional arbitrary shrinkage. A rule-based group average is not inherently uncalibrated either; its quality depends on representative evidence and validation.

Using it, and checking it

Evaluate the model at the age when the decision will actually be made. A six-month forecast may be more informed than a day-seven forecast, but it cannot justify a day-seven acquisition decision. Compare feasible early forecasts against cohort or segment baselines, and include the cost of errors at the resulting budget or targeting thresholds.

As cohorts mature, compare predicted and realized value using the same horizon and cost definition. Inspect errors by channel, region, acquisition period and predicted-value band. A change in acquisition mix or product behavior may degrade the model, but the direction and size are empirical questions. Keep forecasts and their versions so the original predictions can be evaluated without replacing them retrospectively.

Finally, distinguish predicted value from incremental value. A high-pLTV customer may have purchased without an advertisement, and a retention offer may change both revenue and cost. Allocating spend requires evidence about the action’s effect as well as baseline customer value. Use pLTV as an input to that decision, with an explicit uncertainty and evaluation process, rather than equating it with the maximum worthwhile bid.

References

• Fader, Hardie and Lee, Counting Your Customers the Easy Way