How churn prediction works for mobile games
Churn prediction estimates each player's probability of leaving before they do, by learning from your game's own history. Here's how the models actually work, what data they need, and why "explainable" is the part that matters.
Updated July 16, 2026
Churn prediction is a model that estimates, for each player, the probability that they're about to stop playing — before they do. It learns from your game's own history: the behaviour of thousands of past players whose outcome is already known, some of whom stayed and some of whom left. The model finds the behavioural fingerprints that preceded leaving, then scans your current players for the same patterns.
That's the whole idea. No magic — pattern matching against your own past, done more thoroughly than a human could. The interesting parts are which patterns matter, why the model has to wait for outcomes to settle, and what it takes for a prediction to be something you can act on rather than just a scary number.
What signals actually predict churn?
The strongest predictors are almost always behavioural, and they're surprisingly consistent across games:
- Session cadence collapsing. The single biggest tell. A player who played daily and now shows up every third day is drifting, and the drift is visible well before the last session.
- Session length shrinking. Same signal, finer grain — sessions getting shorter while frequency holds is the "checking in out of habit" phase.
- Friction piling up. Failed levels, losing streaks, dead progression. What counts as friction is genre-specific — a fail-and-retry wall in a match-3 game, a losing streak in a sports sim, an offline raid in a survival game. This is why we tag every game with a progression model: so the model knows which events are friction for your genre.
- Progression stalling. Days pass, but level, power, or milestones don't move. The player is logging in but no longer going anywhere.
- The early-behaviour fork. For brand-new players, what happens in the first sessions is hugely predictive — missing the first win or first reward marks a player as at-risk within hours of install. (We wrote a whole guide on that: Why F2P games lose most players on day one.)
Notice what's not on the list: demographics, device model, acquisition channel. They correlate a little. Behaviour dominates.
How do the models actually work?
Under the hood, churn models are usually one of two shapes, and good systems combine them:
Classification asks: "given this player's recent behaviour, what's the probability they won't return within N days?" It's trained on historical players labelled stayed or left, using gradient-boosted trees — a workhorse ML technique that handles messy, mixed-type game data well and doesn't need deep-learning volumes of players to train.
Survival analysis asks a subtler question: "how long will this player keep playing?" Its superpower is handling players whose story isn't over — a player who's been around 20 days and is still active isn't a "stayed" label, they're an unfinished sentence. Survival models (the same family hospitals use for time-to-event analysis) treat them correctly instead of throwing them away, which matters enormously in games where most of your data is current players.
You don't need to remember any of that. What's worth remembering: churn models learn from your game's players — a model trained on someone else's game imports someone else's assumptions.
Why do you have to wait ~14 days for predictions?
Because the model trains on players whose outcome has actually settled. To label a historical player as "churned," we have to know they genuinely left — which means waiting long enough (around 14 days of absence, in most mobile contexts) to be confident they're not just on holiday. A model trained on unsettled labels learns to call every quiet weekend a churn event.
So for a brand-new game starting from an empty dataset, meaningful churn predictions are a matter of weeks, not days — the data physically doesn't exist yet. Anyone who promises accurate churn predictions in week one of a fresh integration is selling you something.
Unless you already have history. If your game has been live on another analytics stack, importing your historical data gives the model years of settled outcomes on day one — which is why it's the single biggest lever on how fast predictions become useful.
How much data do you need?
Honest thresholds, from what we see in practice:
| History | What it unlocks |
|---|---|
| Under 30 days | Nothing predictive yet — descriptive dashboards only |
| 30–89 days | Churn and survival models can train; treat early predictions as directional |
| 90+ days | Solid ground — enough settled cohorts for predictions worth acting on |
Player volume matters too, but less than people fear: the model needs enough settled players to learn from, not millions. A few hundred DAU with months of history beats ten thousand DAU with two weeks.
What does "explainable" mean, and why should you care?
A bare churn score is a number that makes you anxious: "Player X: 73%." Seventy-three percent because of what?
Explainability techniques (SHAP is the standard one) decompose each prediction into its causes: for every input the model used, how much did it push this player's risk up or down? That turns "73%" into "73%, mostly because session length halved this week and the last three levels were all failures."
That difference is the entire practical value of churn prediction. The score tells you who; the explanation tells you what to do — a difficulty tweak for the player stuck on a wall, a comeback reward for the one whose cadence is slipping. Without the reason, a churn model is a weather forecast. With it, it's a to-do list.
What do you actually do with a churn prediction?
The boring, effective playbook:
- Segment by reason, not by score. All the players churning on the same level wall are one design fix, not fifty individual interventions.
- Fix the systemic causes first. If the model keeps naming the same friction event, that's a game problem wearing player-level clothing.
- Intervene while they're still in the game. A hint, an easier path, a well-timed reward — small nudges at the predicted-risk moment, not win-back emails a month after they've gone.
- Watch whether the intervention moves the curve. Predicted-risk cohorts give you a before/after measurement for free.
Frequently asked questions
How much data does churn prediction need?
About 30 days of events before training is possible at all, plus enough players with settled outcomes (~14 days old). 90 days of history is a sensible minimum for predictions worth acting on — importing history gets you there immediately.
Can churn prediction work for a small game with a few hundred DAU?
Yes, with honest caveats: hundreds of DAU and a few months of history is workable; 30 DAU and two weeks isn't. Genre-informed defaults narrow what the model must learn from scratch, which helps most at small scale.
Do I need spend data?
No — the strongest signals are behavioural. Spend events sharpen revenue-weighted predictions (a lapsing whale matters more than a lapsing lurker), but churn models work fine on games with no monetisation.
How accurate is churn prediction?
It depends on your data, and anyone quoting a universal number is marketing to you. Measure it on your own game: train on older cohorts, predict on newer ones, compare against what actually happened.
A note on what we're building: PlayGenus does everything on this page — survival and churn models trained on your game, with a plain-English reason and a recommended action attached to every prediction. You can see it working on demo data at app.playgenus.com/auth/demo, no signup. We're onboarding a small founding cohort of live studios for free — email us if that's you.
Why F2P games lose most players on day one — and what to fix first
Even healthy mobile games lose 60–70% of new players within 24 hours. Day-one churn comes down to a few specific moments in the first session — here's where players actually leave, and how to find your version of it.
What is a good D7 retention rate for mobile games?
The median mobile game keeps under 4% of players at day 7; top-quartile games reach 7–8%, and genre leaders hit the low teens. What counts as "good" depends on which population you compare against — here are the real numbers, with sources.