User churn is easier to prevent than to win back. The problem is that by the time a user stops logging in, it's already too late: they made the decision days ago. We help implement churn prediction — a system that identifies users on the verge of leaving 7–14 days before actual churn, while retention mechanics still work. Over 5 years, we've delivered more than 50 projects in mobile analytics and ML. Our experience shows that timely churn risk detection boosts retention by 15–25% and saves up to 30% of the marketing budget on new user acquisition, translating to savings of $25,000 per month for a typical app with 100k monthly active users.
How We Predict Churn and on What Data
The definition of "churn" depends on the app type. For a daily tracker — not opened for 7 days. For e-commerce — no purchase made in 30 days. For subscription services — cancellation or non-renewal. The model must know this definition upfront.
Features that work in practice:
- Session frequency over the last 7/14/30 days with trend (increasing/decreasing)
- Average session duration and its dynamics
- Number of completed key actions (onboarding steps finished, payment made)
- Days since last session — the strongest single feature
- Progress in the core flow: a user who hasn't added their first diary entry will churn with 80% probability
- Push notification open rate over 14 days
- App version and platform (crashes on a specific version sometimes cause abnormal churn anomalies)
We pull data from mobile analytics: Firebase Analytics, Amplitude, Mixpanel, or a custom event pipeline. The key is proper event setup on the client before starting ML work. Without session_start, key_action_complete, payment_initiated, there's nothing to build the model from.
Why Gradient Boosting Beats Neural Networks for This Task
Gradient Boosting performs best on tabular data with "behavioral" features. XGBoost or LightGBM are the industry standard. Neural networks are overkill here — you likely have a few dozen features, not thousands. XGBoost is 2x more accurate than Logistic Regression and 10% more accurate than Neural Networks on this task.
Typical accuracy on well-prepared data: precision 0.70–0.80, recall 0.65–0.75 at threshold 0.5. Important: optimize for recall, not precision — it's better to send a retention offer to a user who wouldn't have churned than to miss a real churner. Research shows XGBoost achieves up to 80% accuracy on similar tasks.
Training is on historical data with labels: at time T, the user was in the risk group, and after 14 days they actually churned (Y=1) or stayed (Y=0). The class is imbalanced: churners are typically 10–25% of the base. We apply SMOTE or class_weight='balanced'.
How Often Should the Model Be Retrained?
We recommend monthly retraining on fresh data with retrospective labeling. After major product changes, retrain immediately. The process is automated in a pipeline, ensuring predictions stay accurate at 75–80%.
Precision, recall, F1-score, AUC-ROC — the standard set for evaluating model quality. The decision threshold is chosen individually: raising recall increases false positives, raising retention campaign costs. We recommend fixing the threshold after an A/B test.
Which Retention Actions Are Most Effective?
The prediction result is used on the client via Backend-Driven UI or push campaigns:
- Push notifications: for the high-risk segment — a personalized reminder of the app's value. Not "We miss you!" — that doesn't work. Instead: "You haven't logged expenses for 5 days — your budget may exceed the limit." A specific, relevant reason to return.
- In-app messages: on next open — a special offer or an onboarding hint for a user stuck at a certain step.
- Downgrade prevention: if the user visited the subscription settings — trigger a retention offer before cancellation.
Mobile-side integration: at session start, the app requests a config from the backend (Firebase Remote Config or a custom endpoint), receives retention_variant for the current user, and renders the corresponding UI.
What's Included in the Deliverable?
- Documentation on the feature pipeline and churn definition
- Trained model with code and configs
- Integration of batch scoring into your backend
- Setup of retention triggers (push, in-app, remote config)
- Dashboard for monitoring precision/recall
- Team training on working with the system
Backend Infrastructure
User scoring is a batch process, not real-time. We run it daily: pull events from the analytics pipeline (BigQuery, ClickHouse, or your own storage), build a feature vector for each active user over the last 30 days, run through the model, and write to table user_churn_score(user_id, score, risk_segment, calculated_at).
| Segment | Score | Action |
|---|---|---|
| Low risk | < 0.3 | No action |
| Medium risk | 0.3–0.6 | Push notification |
| High risk | > 0.6 | Personalized retention offer |
Algorithm Comparison for Churn Prediction
| Algorithm | Typical Accuracy | Training Speed | Interpretability |
|---|---|---|---|
| XGBoost | 75–80% | High | Medium |
| LightGBM | 73–78% | Very high | Medium |
| Logistic Regression | 65–70% | High | High |
| Neural Network | 70–75% | Low | Low |
Gradient Boosting remains the best choice for mobile analytics due to its balance of accuracy and performance.
How to Implement Churn Prediction: Step-by-Step Plan
- Audit current analytics — check which events are already being sent, whether
session_startand key actions exist. - Define churn definition — fix what counts as churn for your product.
- Design the feature pipeline — create a pipeline for daily feature calculation.
- Collect and label historical data — gather data for the last 6+ months, label churn.
- Train and validate the model — train XGBoost/LightGBM, tune the threshold.
- Integrate scoring into the backend — launch the batch process.
- Set up retention triggers — link scoring with push, in-app, remote config.
- A/B test retention actions — verify model and mechanic effectiveness.
- Monitor precision/recall in production — enable automatic retraining.
An A/B test is mandatory: control group of high-risk users without retention actions, experimental group with them. Otherwise, you won't know if the model works.
Timeline Estimates
Basic model with batch scoring and push notifications — 3–4 weeks if you have 6+ months of historical data. Full system with feature pipeline, A/B test, monitoring dashboard, and automatic retraining — 8–12 weeks. Cost is calculated individually.
Order an audit of your current analytics, and we will propose the optimal solution for your app. Get a consultation on churn prediction setup.