LTV forecasting: cohort curves vs per-player models

Ask two studios how they forecast lifetime value and you'll often get two completely different answers that both sound reasonable. One extrapolates a cohort's spend curve forward from its first weeks of data. The other predicts an individual player's future value from their behavior so far. Neither approach is wrong, but they answer different questions, and using the wrong one for the decision in front of you produces confident numbers that point you the wrong way.

Cohort curves answer a portfolio question

A cohort LTV model looks at everyone who installed in a given week and fits a curve to how their cumulative spend grows over time — steep in the first days, flattening as the cohort ages. Once you have enough historical cohorts, that curve shape becomes predictable enough to extrapolate a new cohort's early spend into a full lifetime estimate before the cohort has actually lived that long.

This is the right tool when the decision you're making is about the portfolio, not the individual: how much can you afford to spend acquiring this week's cohort, is this month's UA channel producing players worth what you're paying for them, is the game's overall economics trending up or down. Cohort curves are stable, easy to audit, and don't require player-level modeling infrastructure — a spreadsheet and enough historical data can produce a usable curve. Their weakness is that they say nothing about any individual player. A cohort LTV of a certain value tells you the average outcome across a group that almost certainly contains players worth far more and far less than that average.

Per-player models answer a targeting question

A per-player LTV model takes an individual's early behavior — session patterns, early spend, progression pace, engagement signals — and predicts that specific player's future value. This is the right tool when the decision is about the individual: which players should see a retention offer worth investing in, which players are worth a personalized win-back campaign, which new players are showing early whale signals worth nurturing differently from the rest of the cohort.

The tradeoff is that per-player prediction is noisier and needs more infrastructure. A cohort curve is fitting one line to thousands of data points; a per-player model is trying to say something meaningful about one player from a handful of early signals, which is a much harder statistical problem. Confidence in an individual prediction should always be lower than confidence in a cohort average, and treating a per-player LTV score as precise when the underlying signal is thin is a common way teams overtrust these models.

The mistake is using one where you need the other

The failure mode isn't picking cohort curves or per-player models — it's using cohort output to make an individual-level decision, or per-player output to make a portfolio-level decision. A studio setting UA bid caps off a per-player model built on thin early signal is bidding on noise dressed up as precision. A studio deciding which specific players get a high-value retention offer based only on a cohort average is treating every player in that cohort identically, which defeats the point of targeting at all.

The two approaches aren't competitors for the same job. A healthy analytics stack runs both: cohort curves for the acquisition-spend and portfolio-health questions, per-player models for the targeting and segmentation questions, and a shared understanding across the team of which one is answering which question at any given moment.

Where the models actually get their signal

Both approaches degrade the same way if the underlying event data is thin. A per-player model with only session-count and days-since-install to work with is guessing with slightly better priors than a coin flip. A per-player model that also has spend timing, progression depth, and response to prior offers has real signal to work with — the difference between those two states is entirely about what your telemetry captures, not about the modeling technique.

This is the practical starting point for a studio unsure which model to trust: audit what signal actually reaches your feature store before debating model architecture. A sophisticated model on thin data underperforms a simple model on rich data almost every time.

Gamemantra's offer pipeline reads directly from your player behavior signals to build these predictions — see how the platform's measurement works, or book a walkthrough to see cohort and per-player forecasting applied to your own data.