Predicted lifetime value, or PLTV, is the modelled value of a player over a chosen horizon. In Sellrise the horizon is 90 days and the answer is always a range, never a single number. The reason is simple: a single number on a 90-day horizon is fiction, and a system that gives you one is hiding its uncertainty. This article explains what PLTV is for, how to read it, and how to avoid the common ways of misusing it.
What does PLTV actually mean?
It is an estimate of how much a player will be worth to you over a period, produced by a model from how that player has behaved so far. Our glossary defines it as the modelled value of a player over a horizon, always shown as a range, never a single number.
It is not what the player has already spent, and it is not a promise. It is a forecast, with all the weaknesses of one. Its value is in comparing players against each other and in deciding where to put effort, not in predicting any one player's exact outcome.
A useful way to think about it: PLTV answers the question of where to look first, not the question of what will happen. If your team has an hour a day for personal contact, the forecast tells them which players to spend it on. That is a modest claim, and it is the one I am prepared to defend.
What is the difference between PLTV and plain LTV?
Plain LTV, in most reports, is a historical total or an average over a cohort. It tells you what happened. PLTV is a per-player forward estimate. It matters most for new players, who have little history, and for players whose behaviour has begun to change.
A cohort average is still useful for planning. In Sellrise the business forecast shows churn and lifetime value by cohort and segment, with an alert when reality drifts from the model. The per-player PLTV is what lets you act on one person today.
Why should the answer be a range?
Because the model does not know the future, it knows the distribution of futures similar players had. A range communicates how much room there is. A player with a narrow range on a strong signal is a good target for a call. A player with a wide range is one to watch.
A single figure invites false precision. Managers will rank by it, set targets on it and argue about the third digit. A range keeps the conversation at the right level. I would treat any vendor who shows one tidy number per player on a 90-day horizon with caution, and ask how the uncertainty is handled.
How is PLTV used in retention?
Three ways stand out:
- Ranking the base. You sort players by value and probability of movement, so the retention budget goes where it returns. In Sellrise this is the player value view: who to grow, who to hold, who to win back, and who to let go.
- Choosing the effort. A call is expensive, so it goes to players whose value justifies it. A message by a cheaper channel goes to the rest.
- Finding VIPs early. A rising PLTV in the first week after the first deposit is a flag for a manager. See how to spot future VIP players in their first week.
Combine it with the probability that the player leaves, since the cost of losing a player depends on both. A high-value player with a moderate risk should often come before a low-value player with a high risk. This is the logic of an expected-value worklist, which I describe in how to find players about to leave.
What does PLTV need as input?
The same four tables as the other forecasts: players, deposits, bets and sessions, keyed by Player ID. No name or contact is required for the model. For quality, I would want a reasonable amount of clean history, remove test, service and bot accounts first, and compute features from raw transactions instead of pre-aggregated statistics that might already include information from the future.
If you have fewer than six months of reliable history, treat any PLTV estimate with extra caution. A model cannot learn what a mature player looks like from a base made of newcomers.
How do you know a PLTV model is any good?
- Time-aware validation. Train on earlier weeks and test on later ones, so the model cannot see the future.
- Calibration of the range. If you say most players fall inside a range, check that they really do.
- Shadow period. Look at the forecasts for a few weeks without acting, then compare with what the players did.
- A controlled test. Use a control group, so you measure whether acting on the forecast moved turnover.
The last step is the one that matters for money. A forecast can be accurate and still useless if no action follows from it, or if the action does not change anything.
What are the common misuses?
- Treating a range midpoint as a target.
- Using PLTV to decide the bonus size without any margin model. Your margins and wagering terms belong to you.
- Using it to exclude players from responsible gaming checks. Value must never override a safeguard.
- Forgetting to retrain. Player behaviour drifts, and a model trained on an older mix will degrade quietly.
On retraining, a weekly refresh on a rolling window is a sensible default, and a monthly one at most.
Where can you see it?
The demo shows value ranges next to churn risk on synthetic data, and the player retention page describes how the budget follows value. To check it on your own base, start a pilot. I would not ask you to trust the range until you have seen it compared with what your players did.