AI and agents
Build a churn prediction model on DataNature, step by step
To build a churn model on DataNature, define churn and the prediction window, build a point-in-time training table from Gold customer and behaviour tables, register the features in the feature store, train and compare models with experiment tracking, register and approve the best one, score customers weekly with a scheduled job, and send the highest-risk customers to retention teams through AI360 Studio, reverse ETL and Voir. Real-time signals from StreamHouse can trigger save actions within minutes.
Published . Last updated
On this page
Churn models are often the first prediction project an organization ships, because the value is easy to see: keep more customers. They also fail often, usually because churn is poorly defined or the scores never reach the people who can act. This guide shows how to avoid both on DataNature.
Step 1: define churn and the window
| Industry | Typical churn definition | Prediction window |
|---|---|---|
| Telecom prepaid | No recharge or usage for 30 to 60 days | Next 30 days |
| Telecom postpaid | Contract cancelled or number ported out | Next 60 to 90 days |
| Banking and credit unions | Primary account closed, or balance and activity drop below a threshold | Next 90 days |
| Insurance | Policy not renewed or lapsed | Before the next renewal |
Agree the definition with the business, and decide what action will follow, such as a call, an offer or a fix.
Step 2: build the training table
Create one row per customer per snapshot date, with features computed from data before that date and the label from the window after it. On DataNature this is a dbt model over Gold tables, so it is versioned and reviewable.
Step 3: engineer features
| Feature group | Examples |
|---|---|
| Activity trend | Transactions, logins, calls or data use in the last 30 days versus the previous 90 |
| Value | Balance or spend trend, products or plans held, tenure |
| Experience | Complaints, failed payments, dropped calls, outage exposure, service contacts |
| Engagement | App use, response to past campaigns, channel preference |
| Life events | Salary moves, address changes, contract end dates |
Register the features in the feature store so the scoring job uses exactly the same definitions. For telecom, network experience features from StreamHouse are often the strongest signals; see network-aware churn save calls.
Step 4: train and compare models
Start with a simple baseline, such as logistic regression, then try gradient-boosted trees. Split by time, track every run in experiment tracking and compare on the metric that matches the action:
- Lift in the top decile or top 5%, if retention teams can only call a fixed number of customers.
- Precision at the chosen threshold, if offers are expensive.
- Calibration, if the probability itself drives decisions or budgets.
Step 5: explain and validate
Store the main drivers of each score, such as "falling balance" or "three complaints this quarter", so staff know what to talk about. Check performance by segment, region and channel before approval, and record the approval in the model registry.
Step 6: deploy and act
- A weekly job in Jobs and Pipelines scores every customer and writes
churn_score,top_reasonsandmodel_versionto a Gold table. - AI360 Studio turns scores into segments and next-best save actions, within offer budgets and consent rules.
- Reverse ETL syncs lists to CRM and campaign tools; Voir briefs relationship managers and care agents; voice agents can call customers who consented.
- StreamHouse triggers real-time actions when a high-risk event occurs, such as a port-out inquiry.
Step 7: measure and monitor
Hold out a random control group, so you measure customers saved by the action, not just customers predicted to churn. Track drift and retrain when performance drops or the business changes.
Frequently asked questions
How much history do we need?
Usually 12 to 24 months, enough to see several cycles and seasons.
Should we predict churn or the effect of an offer?
Start with churn risk. Once you have campaign history with control groups, uplift models can predict who an offer will actually change.
Sources
- DataNature: how to build predictive models
- Use case: subscriber churn prediction
- Use case: customer churn and attrition in banking
How does DataNature Atlas keep this secure and compliant?
Agents are governed like users: scoped, masked, logged and unable to act without approval. DataNature Atlas is the platform's governance module:
- Knows your data: a catalog with automated PII detection and classification across Bronze, Silver and Gold.
- Controls access: role-based grants plus attribute-based policies, row filters and column masks (null, redact, partial, hash, year only).
- Governs AI: Voir agents are scoped principals; masks, row filters and PII guardrails apply to every tool call and answer.
- Proves it: lineage from source to metric to decision, and an audit log of who or what accessed which data.
- Stays local: Atlas runs inside your deployment, so policies, lineage and logs never leave your jurisdiction.
These controls support GDPR, PIPEDA, Quebec Law 25, CCPA, HIPAA and sector rules such as OSFI E-23. Read DataNature Atlas in depth or see security and compliance.
Platform components in this article
Learn more about each DataNature product and module mentioned here.
Products
See it on your own data.
Book a demoDemos and quotes are handled by Dlytica, the company behind DataNature.