You lose 20% of customers monthly and don't know why. Marketing offers discounts to everyone, but the budget grows, and churn doesn't. The standard retention approach in B2B SaaS often fails because it doesn't account for temporal dynamics and root causes of churn.
We encountered this problem in a project for a fintech platform with 200,000 users. Implementing our churn prediction system with personalized recommendations reduced churn by 17% in just 2 months.
The key differentiator is interpretability: you know not only who will leave, but also why, and what to do. Traditional A/B tests and retrospective analysis do not provide an operational picture. High churn undermines LTV and increases CAC.
Our system uses an ensemble of models: Survival Analysis to estimate time to churn and Gradient Boosting for churn probability in a specific window. SHAP values reveal the reasons for each risk, and the recommendation engine generates personalized actions. The result is a 15% increase in retention and up to 30% savings in retention budget.
Problems we solve
Lack of interpretability. Conventional models provide only churn probability but do not answer "why." Without this, targeted action is impossible. Ignoring temporal dependencies. Churn is not an instantaneous event; the customer churns gradually.
Without accounting for time, the model predicts worse. Generic recommendations. All customers receive the same discounts — this is ineffective. Our approach provides a different prescription for each risk profile. Personalized recommendations work 2-3 times more effectively than universal discounts.
How we do it
Stack:
- Survival Analysis: Cox Proportional Hazards and Kaplan-Meier — to estimate time to churn. The accuracy of predicting "when" is higher than classification "will churn/won't."
- Gradient Boosting: XGBoost/LightGBM — for churn probability over 30/60/90 days. SHAP values break down the contribution of each feature: "risk is high due to a 45% drop in usage, 2 open tickets, NPS=5."
- Feature Engineering: usage patterns (daily/weekly/monthly active), feature adoption breadth, login recency, support interactions, email engagement, invoice status.
- Recommendation Engine: for each at-risk customer — personalized actions based on SHAP factors, segment (Enterprise → exec call; SMB → email script), interaction history, and successful cases of similar customers.
Case study. For a fintech platform with 200K customers, we implemented the system in 7 weeks. SHAP revealed that the main cause of churn was a decrease in transaction frequency after the 3rd month.
Recommendation: an automatic personal dashboard and push notifications with spending analysis. Result: churn reduction by 17% in 2 months.
How are churn reasons interpreted?
SHAP attributions are not a black box. For each customer, we get a list of factors with a quantitative estimate of influence.
For example: "the score is elevated due to no login for 5 days (+0.12), low NPS (+0.08), overdue invoice (+0.05)." We convert these data into action recommendations through a rule-based and ML-based model. You can learn more about SHAP in Wikipedia.
Why do personalized recommendations work better?
A/B tests on segments show a retention lift of 10-20% compared to universal discounts. Example: for enterprise customers at high risk, a CSM call increases retention by 25%; for SMB, an email sequence with a case study — by 12%. Without personalization — 0%.
Forrester research shows that personalized communications increase retention by an average of 18%. Investments in the system pay back in 3–4 months due to lower CAC.
Work process
- Data audit — collect logs, CRM, payment history (usually 6-12 months).
- Feature Engineering — build 50+ temporal features.
- Model training — Survival Analysis + Gradient Boosting + hyperparameter tuning.
- Recommendation Engine — rules + ML for selecting actions.
- A/B testing — randomization by customers, 2-week pilot.
- Deployment to production — REST API with latency P99 <100ms.
- Monitoring and retraining — every 2 weeks (retraining on fresh data).
| Stage | Timeline | Result |
|---|---|---|
| Audit and feature engineering | 1-2 weeks | Dataset with features |
| Training and validation | 1-2 weeks | Models with AUC~0.85 metrics |
| Recommendation engine | 1-2 weeks | Action map and API |
| A/B testing and deployment | 2-3 weeks | Ready system with monitoring |
| Approach | Retention lift | Implementation time |
|---|---|---|
| Universal discounts | 0-5% | Week |
| Our system | 15-25% | 7-9 weeks |
Typical implementation mistakes
- Ignoring seasonality — the model overfits to holiday peaks.
- Using outdated data — predicting on data from six months ago is useless.
- No control group — it is impossible to measure lift.
- Generic A/B test — customers are not segmented.
Detailed example of SHAP attribution
For customer A: base value = 0.15 (churn probability); feature contributions: no login for 5 days +0.12, low NPS +0.08, overdue +0.05. Final probability = 0.40. Recommendation: send a push notification with a promo code and call support.What is included in the work
- Model and API documentation (Swagger).
- Integration via REST or WebSocket.
- Grafana dashboard with metrics (predictions, SHAP, retention).
- Team training (2-3 workshops).
- Support for 3 months after deployment.
Evaluate your project in 2 days — write to us. Order an implementation consultation. 5+ years of experience in ML, 50+ completed churn prediction projects, quality guarantee.
