The question.
Predicting who will churn is not the same as knowing whose behavior a campaign can change.
The design separates predictive models from intervention policies and uses experiments to assess whether suggested actions improve customer outcomes.
The proposed approach.
- Build time-aware customer features and cohort definitions
- Calibrate risk and lifetime-value estimates by segment
- Connect recommendations to feasible actions and eligibility rules
- Test actions with holdouts and monitor adverse engagement effects
Target architecture.
Customer eventsFeature snapshotsRisk and value modelsAction policyExperimentCohort reporting
The difficult parts.
- Selection bias in past outreach
- Delayed outcomes and changing customer populations
- Action fatigue and feedback loops
How to evaluate it.
- Calibration and ranking quality on temporal splits
- Incremental retention or value under controlled interventions
- Policy performance across cohorts and contact constraints
Planned deliverables.
- A customer and feature data model
- A calibrated decision-policy design
- An experiment and cohort reporting kit
Design principles.
Predictive risk does not establish treatment effect.
A recommendation must map to an action a team can actually take.