Churn prediction models

Detect which customers are about to leave and act before they do

9 min

Churn (cancellation or attrition rate) is one of the most critical metrics in subscription, SaaS and recurring ecommerce businesses. Acquiring a new customer costs 5 to 7 times more than retaining an existing one, according to Bain & Company. Churn prediction models identify customers with high abandonment probability before they leave, enabling proactive retention strategies.

An effective churn model doesn’t just predict who will leave — it explains why. This explainability is what allows designing specific interventions: a customer leaving over pricing requires a different approach than one leaving due to poor support.

Early warning signals of churn

Customers who will cancel typically emit signals weeks or months beforehand. These signals vary by industry and business model, but patterns are consistent: declining usage, reduced engagement and signs of dissatisfaction.

  • Usage decline: less frequent logins, key features unused, shorter sessions
  • Decreasing engagement: unopened emails, ignored notifications, unconsumed content
  • Dissatisfaction signals: frequent support tickets, negative reviews, low NPS
  • Behaviour changes: plan downgrade, user removal, data export
  • Contractual factors: approaching contract end, change of purchasing contact

Data needed for the model

A churn model needs data describing customer behaviour over time. The more quality data you have, the more accurate the model will be. Data falls into four main categories.

  • Product usage data: login frequency, features used, depth of use, active sessions
  • Interaction data: emails opened/clicked, support contacted, webinars attended, content consumed
  • Account data: tenure, subscribed plan, number of users, upgrade/downgrade history
  • Demographic/firmographic data: company size, sector, region, decision-maker role
  • Satisfaction data: NPS, CSAT, reviews, direct feedback

How to build a churn model

Building a churn model follows the standard machine learning workflow: define the objective, prepare data, select features, train the model, validate and deploy. The churn-specific challenge is correctly defining what "churn" means in your context.

In SaaS, churn may be subscription cancellation. In ecommerce, it might be no purchase in 90 days. In an app, it could be not opening the application in 30 days. The definition must be precise, measurable and aligned with the business. Once defined, the historical dataset is labelled: this customer churned, this one didn’t.

  • Define churn precisely: what action (or inaction) constitutes abandonment and over what time window
  • Build features from raw data: averages, trends, ratios, counts per period
  • Split data into train (70–80%) and test (20–30%) respecting chronology
  • Start with interpretable models (logistic regression, Random Forest) before trying complex ones
  • Evaluate with appropriate metrics: AUC-ROC, precision, recall and lift in top deciles

Prediction-based retention strategies

A churn model without action is merely an analytical exercise. The real value lies in interventions triggered when a customer exceeds a risk threshold. The most effective interventions are personalised according to the probable cause of churn.

  • Proactive outreach: personal contact from the CSM or account manager before the customer complains
  • Retention incentives: discounts, free extensions or temporary upgrades for at-risk customers
  • Improved onboarding: if customers churn due to poor adoption, strengthen initial guidance
  • Pain point resolution: if support is causing churn, prioritise at-risk customers’ tickets
  • Educational content: webinars, guides and tutorials demonstrating the product’s value

Implementation and operationalisation

A churn model must be integrated into the systems the team uses daily. If the model generates a list of at-risk customers that nobody reviews, it adds no value. The most effective integration is directly in the CRM: each account displays its updated churn probability and recommended actions.

Update cadence depends on the business: monthly models work for annual contracts, but businesses with fast churn (apps, monthly subscriptions) need weekly or daily scoring. Automating the data pipeline (extraction, feature engineering, scoring, distribution) is essential for the model to work in production.

Retention programme impact metrics

The model’s success is measured by its business impact, not its technical accuracy. A model with 0.85 AUC that reduces churn by 15% is more valuable than one with 0.95 AUC that nobody uses.

  • Churn rate reduction: before/after comparison following intervention implementation
  • Net Revenue Retention (NRR): retained revenue including expansions and contractions
  • Customer Lifetime Value (CLTV): increase in average customer lifecycle value
  • Retention cost vs acquisition cost: retention programme efficiency
  • Model lift: how much better the prediction is compared to a random baseline

Key Takeaways

  • Customers emit churn signals weeks before cancelling: declining usage, less engagement, dissatisfaction
  • Precisely defining what churn means in your context is the critical first step
  • An interpretable model explaining churn causes is more valuable than one that’s merely accurate
  • Personalised interventions based on churn cause are far more effective than generic ones
  • The model must be integrated into the CRM and updated at the right cadence for the business

Want to predict and reduce customer churn?

We build churn prediction models integrated into your CRM that identify at-risk customers and trigger personalised retention strategies.