Predicted LTV models: what to demand from vendors
The inputs a pLTV model actually needs, why calibration drifts, and the simple test that shows whether a model beats a day-3 revenue ratio.
Every predicted LTV pitch arrives with the same promise: bid on where a player is going, not just where they have already been. The promise is reasonable. Execution varies enormously across vendors, and most of that variance stays invisible from a sales deck. A buyer evaluating a pLTV product has to get specific about inputs, about how often the model recalibrates and about what happens when it drifts, because the marketing language is close to identical across vendors while the underlying model quality is not.
What inputs the model actually needs
A pLTV model earns its keep by using signal a buyer cannot see manually: session depth and timing in the first hours, progression pace relative to your game's own historical cohorts, and, where available, early monetisation events even below a full purchase, currency accrual, ad views, offer views. A model fed only install-time and day-1 metrics is not meaningfully different from a rules-based day-1 ratio dressed up in machine-learning language. Ask the vendor precisely which features feed the model and at what granularity. Ask over what time window each feature gets collected, too. A product advertised as "predicted LTV" that only ingests coarse SKAN-style conversion values is working with far less than the name implies.
For iOS specifically, this matters more than vendors tend to volunteer. You cannot simply repoint a model built for Android's device-level event stream at SKAN's six-bit conversion value and expect equivalent quality. Any vendor claiming otherwise should show you the accuracy gap between the two platforms directly rather than assert there isn't one.
Calibration windows and drift
A model calibrated once at onboarding and left alone degrades as your game changes: a rebalanced economy, a new event type, a shifted acquisition mix, all quietly move the relationship between early signal and eventual revenue that the model learned. Ask how often the vendor recalibrates against your actual outcomes. Ask what the minimum data volume is before recalibration means anything statistically, and whether recalibration happens automatically or waits for you to notice the drift and request it.
A vendor with no stated recalibration cadence is asking you to trust a snapshot indefinitely.
Drift isn't hypothetical. A live-ops event that temporarily inflates early currency spend without changing eventual payer conversion will, for a period, make the model look like it is under-predicting or over-predicting depending on which direction the event skews behaviour, and a model without a drift-monitoring process will keep bidding on the stale relationship well past the point it stopped holding.
The test that actually reveals model quality
The single most useful evaluation a buyer can run before committing spend against pLTV bidding is a direct comparison against the simplest possible baseline: a day-3 revenue ratio, meaning a player's day-3 cumulative revenue multiplied by a fixed historical factor derived from your own past cohorts. Take a holdout set of installs the model has already scored, wait for actual day-30 or day-60 revenue to mature, and compare the pLTV model's predictions against both the actual outcome and the naive day-3 ratio's predictions for the same cohort.
A model earning its cost premium should beat the day-3 ratio by a clear and consistent margin across multiple cohorts, not just in one favourable test window. If it ties with a ratio you could compute in a spreadsheet, or loses to one, it isn't providing the value its pricing implies, whatever the demo showed on a vendor-selected example cohort.
A model that consistently over-predicts on new campaigns but corrects only after several weeks of live data is showing a common failure mode: someone calibrated it on a mature dataset and it struggles with cold-start conditions, exactly when a buyer most needs a reliable early read to make a scale-or-kill decision. Watch for the flat scorer as well. A model that produces near-identical predicted values across visibly different cohorts, a casual puzzle title and a mid-core strategy game scored by the same vendor, is leaning on genre-agnostic averages rather than features specific enough to tell the two apart.
Questions worth putting in writing before signing
Ask for the model's accuracy metrics on a genre comparable to yours, not an aggregate figure blended across every game type the vendor serves, since pLTV accuracy varies sharply between a deep RPG with a long payer tail and a casual game that monetises mostly through ads in the first week. Ask what happens to bidding when the model has too little data for a new campaign or a new country, whether it falls back to a simpler heuristic or keeps predicting confidently on thin evidence. And ask directly for the day-3 ratio comparison described above as part of the proof of concept. A vendor confident in their model will run that test happily. One who resists it is telling you something.
