Glossary
What is predictive analytics?
Predictive analytics uses historical data, statistical methods and machine learning to forecast future user behaviour, such as churn risk, lifetime value or conversion likelihood.
What it is used for
- Predicted [LTV](/resources/glossary/lifetime-value-ltv), so acquisition decisions can be made before the lifetime has elapsed.
- [Churn](/resources/glossary/churn) risk, to intervene before a user leaves rather than after.
- Conversion likelihood, to prioritise effort where it will pay.
- Propensity scoring, for offer and message selection.
- Demand forecasting, for inventory and budget planning.
Why predicted LTV became necessary
Privacy changes shortened the measurement window drastically. On iOS, the conversion value can only be updated for a limited period after install, so the signal available for optimisation covers days rather than months.
That makes early prediction unavoidable rather than optional: you must estimate long-term value from a few days of behaviour, because a few days of behaviour is all the framework will report. Predicted LTV moved from a sophistication to a requirement.
How it fails
Training on the wrong population. A model built on early adopters will overestimate the value of users a scaled campaign brings, and the error appears exactly when budgets are largest.
Distribution shift. A model trained before a pricing change, a product change or a market shift keeps applying relationships that no longer hold, and does not announce it.
Self-fulfilling loops. A model that predicts low value for a segment causes less investment in that segment, which produces the low value the model predicted.
Precision theatre. A prediction reported without its confidence interval invites decisions the model cannot support.
The discipline that catches all four is holdout validation — reserving users the model did not see and checking what actually happened.
Common questions
How much data do I need for predictive LTV?
Enough completed lifetimes to train on, which usually means at least several months of history across a range of acquisition sources. Models built on one source generalise poorly to others.
How often should models be retrained?
On a schedule and after any material change to product, pricing or acquisition mix. Monitoring prediction error over time is what tells you when the schedule is wrong.
More in analytics and data
Where the numbers are stored, shaped and read.