LTV
Probabilistic LTV without waiting a year
Identity Flowcore journal · notes from On-Device Probability
Product wants a lifetime value on day seven. Finance wants a year of observed cash. Both requests are understandable. Neither is entitled to fake precision. Predictive mobile analytics can speak in probabilities early; it should not speak in fake certainty.
What a short window can support
Seven or thirty days of in-app behaviour can support a rank: who is more likely to still be paying at day ninety than their neighbour, given similar acquisition. It can support a distribution: a band, not a point. It cannot support a board slide that says “this cohort is worth THB X for the next twenty-four months” unless you have history that long and a stable product.
We like models that output a survival curve with wide intervals and a written list of missing information: off-app payments, shared devices, refunds that arrive late, agents who complete onboarding for the user.
On-device is not magic
Running a tiny scorer on the phone can reduce latency for in-session offers. It does not create new truth. The features available on-device are a subset, often without warehouse joins, often without refunds. Treat on-device scores as a cache of a server judgement, or as a deliberately weaker cousin, and say so in the operating note.
Sentences we ask teams to refuse
“Predicted LTV” used as a credit rank. “Guaranteed payback in 11 days.” Any claim that ignores refunds and chargebacks. Any comparison of iOS and Android LTV that does not mention store fees and ATT drop-off.
Collections and credit teams are especially hungry for a number that looks like a bureau score. If that is the request, the honest answer is often no. One of our case notes is about drawing that line during Atelier.
If you need the longer critique path, start with the course desk or write to us with the decision the score is supposed to protect.