Table of Content
What is churn prediction?
Churn prediction uses past customer data to estimate which customers are likely to leave, cancel or stop buying in a set window, such as the next 30 or 90 days. It starts with a simple health score built from usage, support, billing and engagement signals, and can grow into a machine learning model. The goal is not the score itself. It is to warn your team early enough to act.
The short version
- Start with signals, not algorithms. A health score built from 9 clear signals often beats a complicated model on messy data.
- Define churn before you model it. Canceled, did not renew, or no purchase in 90 days are different outcomes with different data.
- Accuracy is a trap. If 5% of customers churn, a model that says nobody will churn is 95% accurate and useless. Check precision, recall and lift.
- A score needs a playbook. Every risk band should trigger a named action by a named owner.
- Watch for data leakage. Signals that only appear after a customer decides to leave make a model look brilliant and fail in real use.
- Want a model built on your data? Our predictive analytics services team builds churn and health models that plug into your CRM.
What churn is and how to run a churn analysis
Churn is customers lost in a period. How you define a lost customer depends on how you sell. A subscription business counts cancellations and non-renewals. An online store counts customers who have not bought for longer than their usual gap. A services company counts clients who end or do not renew a contract.
Write the definition down before you build anything, because customer churn prediction is only as good as its label. Then measure it with a few standard formulas:
| Metric | Formula | Use it for |
|---|---|---|
| Customer churn rate | Customers lost in the period / customers at the start of the period, x 100 | How fast you lose customers |
| Revenue churn rate | Recurring revenue lost to cancellations and downgrades / recurring revenue at the start, x 100 | How much revenue you lose |
| Net revenue retention | (Starting recurring revenue + expansion – contraction – churn) / starting recurring revenue, x 100 | Whether growth from existing customers offsets losses |
| Retention rate | 1 – customer churn rate | The mirror view |
| Repeat purchase gap | Median days between orders for a customer segment | Setting the “lapsed” line for stores |
People ask what a good churn rate is. There is no universal answer. It varies by model, price point and customer size, and a monthly rate and an annual rate are not comparable. Compare against your own history and your own segments. For the retention side, see our guide to customer retention strategies.
9 churn prediction signals to build into a health score
A customer health score combines signals into one number or a red, amber and green rating. These nine signals work across most subscription, store and service businesses. Choose the ones you can measure reliably and give each a weight you can explain.
1. Product or service usage
Signal: Logins, active days, features used or orders placed compared with the customer’s own norm.
Why it predicts churn: Usage usually falls before cancellation. The size of the drop matters more than the level.
Automate: A weekly job that compares each customer with their last 8-week average and flags falls beyond a threshold.
Choose it if you can see usage or order data. Watch for comparing small and large customers with the same threshold. Use each customer’s own baseline.
2. Recency of last activity
Signal: Days since the last login, order or meaningful action.
Why it predicts churn: A long gap beyond a customer’s normal rhythm is one of the clearest early warnings.
Automate: Compute each customer’s normal gap, then alert when the current gap passes about twice that.
Choose it if you sell on repeat. Watch for using one cutoff for all. A monthly buyer and an annual buyer differ.
3. Support tickets and complaints
Signal: Open tickets, unresolved age, repeated issues, low satisfaction scores and escalations.
Why it predicts churn: Unresolved problems erode trust. Many tickets can also mean an engaged customer, so combine count with sentiment and resolution time.
Automate: Pull tickets from your help desk and tag sentiment. Our guide to Zendesk alternatives covers help desks with AI support built in.
Choose it if you run a help desk. Watch for treating all tickets as bad. Fast, resolved tickets can build loyalty.
4. Billing and payment signals
Signal: Failed payments, late invoices, downgrades, removed seats and requests to cancel.
Why it predicts churn: Some churn is involuntary, caused by expired cards. That is the easiest churn to fix.
Automate: Retry failed charges, send card update reminders and flag accounts with repeated failures. Our guide to payment reconciliation automation covers matching payments to accounts.
Choose it if you charge recurring fees. Watch for confusing involuntary and voluntary churn. They need different fixes.
5. Engagement with your communication
Signal: Email opens and clicks, replies to account managers, webinar attendance and response to surveys.
Why it predicts churn: Silence is a signal. Customers who stop responding are often disengaging.
Automate: Track engagement per contact, and alert when a key contact stops replying.
Choose it if you have regular touchpoints. Watch for open-rate noise. Privacy features inflate opens, so use clicks and replies as well.
6. Champion or contact changes
Signal: Your main contact leaves, changes role or is no longer in the account’s activity.
Why it predicts churn: In business relationships, a lost champion often ends the renewal.
Automate: Monitor for bounced emails and job-change alerts and flag accounts with a single contact.
Choose it if you sell to businesses. Watch for relying on one contact. Build multi-threaded relationships.
7. Contract and renewal timing
Signal: Days to renewal, contract length, auto-renew status and past renewal behavior.
Why it predicts churn: Risk is concentrated near the end of terms. Time the playbook accordingly.
Automate: Build a renewal calendar with risk score and a 90-day countdown.
Choose it if you have contracts or annual plans. Watch for starting too late. Begin well before the date.
8. Value realized
Signal: Whether the customer has reached the outcome they bought for, such as first project launched, first campaign completed or savings achieved.
Why it predicts churn: Customers who never see value leave, whatever the price.
Automate: Define two or three “value moments,” log when each happens and flag customers who miss them by day 30, 60 or 90.
Choose it if your product has a clear first win. Watch for measuring activity instead of outcomes.
9. Customer fit and profile
Signal: Industry, size, acquisition channel, plan type and tenure.
Why it predicts churn: Some segments churn more. Knowing this helps you set expectations, adjust targeting and weight scores.
Automate: Join CRM data with product and billing data. Our customer data platform service unifies those sources, and our guide to lead scoring models shows a similar scoring approach for the front of the funnel.
Choose it if you have varied customer types. Watch for using fit alone. Combine it with behavior.
How to turn the signals into a health score
| Step | What to do | Example |
|---|---|---|
| 1. Pick signals | Choose 5 to 9 you can measure reliably | Usage drop, days since last order, open tickets, failed payment, days to renewal |
| 2. Normalize | Convert each to a 0 to 100 scale or a flag | Usage down 40% = 20 points |
| 3. Weight | Give more weight to signals that best separate churners | Usage 30%, support 20%, billing 20%, engagement 15%, renewal 15% |
| 4. Band | Set red, amber and green thresholds | Below 40 red, 40 to 70 amber, above 70 green |
| 5. Test | Check how past churners would have scored 60 to 90 days earlier | If most churners were red, the score has signal |
| 6. Tune | Adjust weights every quarter | Remove signals that do not separate groups |
Start simple and keep it explainable. Account managers will use a score they understand. Move to a model when you have enough history and the simple score hits its limits.
How to build a churn prediction model in 8 steps
A churn prediction model learns from past customers who stayed and left, then scores current customers. Here is the practical path:
- Define churn and the window. For example: a customer cancels within the next 90 days.
- Build a snapshot table. One row per customer per snapshot date, with signals as they stood on that date and a label showing whether they churned in the next window.
- Split by time, not at random. Train on older snapshots and test on newer ones, so the test mirrors real use.
- Start with a baseline. A rules-based score or logistic regression shows what is achievable and is easy to explain.
- Try tree-based models. Gradient boosted trees often do well on tabular customer data.
- Evaluate with the right metrics. Use precision, recall, lift and the number of at-risk customers your team can handle, not raw accuracy.
- Explain the drivers. Show the top reasons for each score so a person knows what to do.
- Deploy, monitor and retrain. Score weekly, track drift and compare predicted with actual churn each quarter.
Traps that make a model look better than it is
- Data leakage. A signal such as “cancellation requested” or “account closed” only exists once the customer has decided. Remove anything recorded after the snapshot date.
- Class imbalance. When few customers churn, accuracy misleads. Use lift in the top 10% of scores and the precision at the number you can contact.
- Random splits. They mix future and past data and flatter the model.
- Ignoring action capacity. Scoring 2,000 customers as high risk helps nobody if the team can call 100.
Tools and approaches for churn prediction
| Approach | Best for | Watch for |
|---|---|---|
| Spreadsheet or BI health score | First version, small customer base | Manual refresh, limited signals |
| CRM or customer success platform scoring | Teams already in a CRM such as HubSpot or Salesforce, or in tools such as Gainsight or ChurnZero | Limits on custom data and modeling |
| Python with scikit-learn or gradient boosting libraries | Custom models and full control | Needs data skills and monitoring |
| Warehouse or cloud ML (BigQuery ML, Databricks, SageMaker) | Large data already in a warehouse | Setup and cost |
| Custom build with an AI agent on top | Joining signals, explaining scores and drafting outreach | Governance and human review |
For the dashboard layer, see our guides to executive dashboard examples and financial dashboard examples, which both include retention views. If you are choosing a CRM to hold the scores, our comparison of the best CRM software covers AI features and pricing.
What to do with a risk score: playbooks that work
| Risk band | Trigger | Action | Owner |
|---|---|---|---|
| Red | Score drops into red, or renewal within 90 days with low usage | Call from an account manager, review value, offer help or a plan change | Account manager |
| Amber | Score falling for two periods | Personal email, tips for the unused feature, invite to a training session | Customer success |
| Green, expanding | High score and rising usage | Ask for a review or referral, suggest an upgrade | Account manager |
| Payment risk | Failed charge | Retry, then card update reminders by email and in the product | Billing |
| Lapsed (stores) | Gap beyond about twice the usual | Win-back email or SMS, with a reason to return | Marketing |
Measure whether the playbook works by keeping a holdout group that gets no outreach and comparing churn. Otherwise you cannot tell whether the score or the action made the difference.
Where AI agents fit
A model gives a score. An agent turns it into work. It can read the account history, summarize why the score fell, draft a personal message for the account manager, create a task in the CRM and log the outcome. For stores, it can pick the win-back offer by segment.
Keep a person approving customer messages at first. Our guides to AI sales agents and building an AI agent for your business explain the safeguards, and our AI automation team builds agents that act on your churn scores.
7 churn prediction mistakes to avoid
- No clear churn definition. Teams argue about numbers because the label differs.
- Starting with a complex model. Ship a simple score first and learn from it.
- Using accuracy as the headline. Imbalanced data makes it meaningless.
- Leaking future data. The model then fails when it goes live.
- Scoring without an action plan. A red score with no owner changes nothing.
- Never retraining. Customer behavior and products change.
- Ignoring involuntary churn. Failed payments are often the cheapest fix.
Frequently Asked Questions
What is churn prediction?
Churn prediction estimates which customers are likely to leave or stop buying in a future window, using signals such as usage, support, billing and engagement. It can be a simple health score or a machine learning model, and it exists to trigger early action.
How do you calculate churn rate?
Divide the number of customers lost in a period by the number of customers at the start of the period, then multiply by 100. Revenue churn uses recurring revenue lost instead of customer counts. Use the same period and definition every time.
What is a good churn rate?
There is no single good rate. It depends on your model, price and customer size, and monthly and annual rates differ. Compare your rate with your own history and your own segments, and aim for a steady decline.
What is a customer health score?
A customer health score combines signals such as usage, support, payments, engagement and renewal timing into one number or a red, amber and green rating. It shows which customers need attention, and it should be explainable to the team that uses it.
How accurate are churn prediction models?
It depends on data quality, history and how you define churn. Accuracy is a poor measure when few customers churn. Judge a model by precision, recall and lift at the number of customers your team can contact.
What data do I need for churn prediction?
You need a clear churn label, customer attributes and signals over time, such as usage, orders, tickets, billing events and communication. Usually a year or more of history helps, with snapshots taken at regular dates.
Which machine learning model is best for churn?
Logistic regression is a good, explainable baseline. Gradient boosted trees often perform better on tabular data. The best choice is the one that performs well on a time-based test and that your team can explain and maintain.
What is the difference between voluntary and involuntary churn?
Voluntary churn happens when a customer decides to leave. Involuntary churn happens when payment fails or a card expires. Involuntary churn is often the easiest to reduce with retries and reminders.
Can small businesses use churn prediction?
Yes. A small business can start with a simple score in a spreadsheet or CRM using recency, orders and support signals. Move to a model when you have enough customers and history to learn from.
How does AI help reduce churn?
AI can combine signals into scores, explain why a customer is at risk, draft outreach and create tasks for the right owner. A person should review customer-facing messages until the process proves reliable.
Bottom Line
Churn prediction works when it is simple, explainable and tied to action. Define churn, pick the signals you trust, build a health score, test it on past customers and give each risk band a playbook. Move to a machine learning model when the data and the team are ready.
If you want a health score or model built on your data, explore our predictive analytics services or our AI automation team.
