Predicting LTV from Player Behavior Patterns

Ilya Pakharukov1 Nov 2024

Problem

A free-to-play publisher pays for installs before it knows who will play. On PC an install costs 5 to 10 USD in the speaker's words (slide 7: 6 to 10-15 USD), and only 2-5% of newcomers ever pay. Payback may take one to three years, yet the traffic budget is spent now. Marketing asks whether the traffic will pay back, and almost nothing is known about new players in advance.

Example from the article: in one Astrum Entertainment game, a model assigns each new player to one of five behavior clusters from early play, and the cohort's LTV is then read off the clusters.

Solution in brief

Instead of predicting money directly, the authors predict which behavior pattern a new player belongs to. The patterns come from a 2x2 matrix ("likes the game" by "will pay") plus an origin cluster. Each pattern has a stable LTV curve over years of cohorts, so cohort LTV is the sum of players per cluster times the cluster's LTV. This gives a fast scoring forecast, not an exact one.

Questions this case answers
  • How to forecast the lifetime value of new users within days from early behavior when payments come years later and only a tiny share of users ever pay
  • How to decide whether acquisition traffic will pay back by assigning newcomers to behavior segments that have known long-run value curves
  • How to estimate the payback of purchased traffic without waiting months for payments
  • How to predict LTV of a new player from early in-game behavior

Tools and methods

  • Behavioral clustering — finds clusters that fit the matrix.
  • ML model on early behavior — assigns a new player to a cluster.
  • Cluster-level LTV curves — convert cluster counts into money.
  • A/B tests on mechanics — check whether a change moves players between clusters.
Continue reading — it's free

Google or a one-time email code. No password.