SaaS churn prediction gets all the attention. App uninstall prediction is the harder, more urgent problem. Harder because the signals are noisier (mobile usage is inherently spiky). More urgent because the decision-to-delete window is shorter and the consequences are more final.
But the fundamentals are the same. You collect behavioral data, train a model to find patterns that precede uninstalls, and use those predictions to trigger interventions. The difference is in which signals matter and how fast you need to act.
Why uninstall prediction needs its own model
You can't just port a SaaS churn prediction model to mobile. The dynamics are different:
- Shorter decision window. SaaS customers think about canceling for weeks. App users decide to uninstall in seconds, usually during a storage cleanup or after a frustrating experience. Your model needs to flag risk 7-14 days early, not 30-60 days.
- Noisier signals. A SaaS user who hasn't logged in for 5 days is concerning. A mobile user who hasn't opened your app in 5 days might just be on vacation. Mobile usage patterns are more variable, so your model needs to account for personal baselines.
- Fewer intervention channels. If a user has already disabled push notifications, your only real-time channel is gone before the uninstall happens. You need to act while you still have push access.
- No explicit cancellation. SaaS users click a cancel button, giving you a save flow moment. App users just long-press and delete. There's no cancel page to intercept.
The signals that predict uninstalls
I've seen these features consistently work in uninstall prediction models, ranked by importance:
Tier 1: Strong predictors
- Session frequency trend. Not the absolute count, the rate of change. A user going from 5 sessions/week to 2 sessions/week is at higher risk than a user who's always been at 2/week. Calculate the 7-day moving average and compare to the user's 30-day baseline.
- Days since last session. Simple but powerful. Every day without a session increases uninstall probability. The curve is exponential, not linear: the jump from day 5 to day 7 is more significant than day 1 to day 3.
- Notification interaction rate. Track the ratio of notifications tapped vs. dismissed vs. ignored over the last 14 days. A declining tap rate is one of the earliest warning signs because it means the user is mentally disengaging before they physically uninstall.
- Push token status. If a user explicitly disables notifications in their OS settings, uninstall probability jumps 3-4x. This is your last-resort signal.
Tier 2: Moderate predictors
- Session duration trend. Sessions getting shorter means the user is getting less value each time they open the app.
- Feature usage breadth. Users who only use one surface-level feature are at higher risk than those who use multiple features deeply.
- App update behavior. Users who skip app updates (still on an old version) are less invested.
- Permission changes. Revoking previously granted permissions (location, camera) signals declining trust.
Tier 3: Contextual signals
- Device storage level (if accessible via your SDK). Low storage triggers OS cleanup prompts.
- OS version. Major OS updates sometimes change notification settings or permissions, causing unintended disengagement.
- Install source. Users from paid campaigns uninstall at higher rates than organic installs.
- Day of week and time patterns. Weekend-only users have different risk profiles than daily users.
Building the model
The same approach from the SaaS churn prediction guide applies, with adjustments:
Training data: you need 3+ months of user data with clear uninstall events. Most analytics platforms (Firebase, Amplitude, Adjust) track uninstall events through push token invalidation or silent push checks. Make sure your tracking is reliable, because uninstall detection has a 24-48 hour lag on most platforms.
Algorithm: gradient boosted trees (XGBoost/LightGBM) work well here, same as SaaS prediction. The dataset is usually larger (more users than a typical SaaS), which helps accuracy. Aim for AUC-ROC of 0.75+ to make automated interventions worthwhile.
Prediction window: train the model to predict "will this user uninstall within the next 14 days?" That gives you enough runway for a 2-3 touch intervention sequence. Shorter windows (7 days) are more accurate but give you less time to act.
Scoring cadence: run predictions daily for all active users. Unlike SaaS where weekly is fine, mobile behavior changes faster and you need to catch shifts early.
The intervention framework
Predictions without interventions are useless. Here's the tiered system that works:
Tier 1: Low risk (score 0.3-0.5)
User is showing early signs of disengagement but hasn't committed to leaving.
Intervention: subtle value reinforcement. A push notification that surfaces something personally relevant: "Your team shared 3 new updates" or "You have 12 unread insights from this week." The notification should deliver value, not ask for attention. No "We miss you!" messages. Those accelerate uninstalls.
Tier 2: Medium risk (score 0.5-0.7)
User is clearly pulling away. Sessions declining, notifications being dismissed.
Intervention: a more direct in-app message on their next session (if there is one) plus a targeted push. Show them a feature they haven't tried that's popular with similar users. Or surface their own historical data: "You logged 47 workouts this year. Keep your streak going?"
If they have email on file, consider sending a helpful email at this stage too. The AI retention email guide covers how to personalize these.
Tier 3: High risk (score 0.7+)
User is very likely to uninstall soon. Push notifications may already be disabled.
Intervention: email is your primary channel now. Send something that acknowledges value they've built: data they'd lose, progress they've made, connections they've built in the app. Make the re-engagement reason specific and personal, not generic.
If they do come back, make the return experience frictionless. Don't show onboarding again. Surface exactly where they left off.
After the uninstall: offboarding recovery
Some users will uninstall despite your best interventions. When that happens, email is all you have left. A proper offboarding email sequence recovers 2-5% of uninstalled users. That sounds low, but when you're losing thousands of users per month, 3% recovery is meaningful revenue.
The key is detecting the uninstall quickly (within 24-48 hours) and triggering the email sequence immediately while the app is still fresh in the user's memory. Most analytics platforms can detect uninstalls through push token invalidation and send webhook events you can connect to your email system.
If your lifecycle emails have gaps you haven't audited recently (most do), a Retention Email Snapshot can surface where you're losing users to email silence.
For the full uninstall prevention experiment framework, see the uninstall reduction experiment. And for context on how this fits into the broader AI retention picture, read the AI churn reduction guide.