Churn Prediction Models for Marketing Retention: Andrew Luxem

The gap between prediction and prevention

Every marketing team I've worked with in the last five years has either built a churn model or talked about building one. Most of the models that get built share a common fate: they produce a score, the score goes into a report, the report gets reviewed in a weekly meeting, and nothing operationally changes.

The model isn't the problem. The problem is the last mile. A churn score that doesn't trigger an action in your ESP, your ad platform, or your retention team's workflow is an expensive monitoring tool. Monitoring is not retention.

The teams that actually reduce churn treat prediction as an input to a system, not the output of a project.

Defining churn is harder than it sounds

Before you build anything, you need a churn definition that matches your business model. This step gets skipped more often than it should.

For subscription businesses, churn has a clear event: cancellation. Even there, you have choices. Is a customer who pauses for 60 days churned? What about someone who downgrades to a free tier? Your model's predictions are only as useful as the definition they're targeting.

For non-contractual businesses (retail, e-commerce, marketplace), there's no cancellation event. Churn is inferred from inactivity. A customer who hasn't purchased in 90 days might be churned, or they might buy seasonally. The definition needs to reflect actual purchase cycles, not arbitrary time windows.

At Ancestry, this distinction mattered constantly. A subscriber who cancelled was clearly churned. But a DNA-kit buyer who hadn't purchased again in six months wasn't necessarily lost: the product has a natural one-time purchase pattern. Lumping both into the same churn definition would have produced a model that was technically accurate and practically useless.

Get the definition right first. The model follows.

The signals that actually predict churn

Feature engineering is where marketing domain knowledge matters more than data science technique. The best churn models aren't the ones with the most sophisticated algorithms. They're the ones with the best input features.

Engagement decay rate. Not whether engagement is low, but how fast it's declining. A customer who went from opening every email to opening one in four over the past month is a stronger churn signal than a customer who's always opened one in four. The trajectory matters more than the level.

Support contact patterns. A customer who contacts support once about a product issue and gets it resolved is fine. A customer who contacts support three times in 30 days is at risk. Repeated support contacts are one of the strongest churn predictors across every business I've worked in, and most marketing teams don't have access to this data. Get it. It's worth the integration work.

Purchase interval stretching. If a customer's average time between purchases is 25 days and their current gap is 40 days, that's a signal. Not because 40 days is inherently long, but because it's long relative to their pattern. Personalized baselines beat population averages.

Feature or category contraction. A customer who used to buy across three categories and is now buying in one is narrowing their relationship with you. That contraction often precedes full churn by 60 to 90 days. It's a signal most teams miss because they're looking at frequency, not breadth.

Payment failures and billing friction. For subscription businesses, involuntary churn from failed payments can represent 20-40% of total churn. This isn't a prediction problem: it's an operations problem. But your model should distinguish between voluntary and involuntary churn because the interventions are completely different.

Building a model marketing teams can use

The technical bar for a useful churn model is lower than most teams assume. You don't need a neural network. You need clean features, a sensible target variable, and a deployment path.

Logistic regression or gradient-boosted trees (XGBoost, LightGBM) handle the majority of churn prediction use cases well. They're interpretable, they train fast, and they produce probability scores that map cleanly to marketing actions.

The output should be a probability score between 0 and 1, refreshed at a cadence that matches your intervention timeline. If your retention flows take a week to activate, a daily score refresh is sufficient. If you're triggering real-time interventions, you need a pipeline that scores on event.

Here's what matters more than model selection: the score needs to land somewhere actionable. That means pushing it into your customer data platform, your ESP audience builder, or your CRM as a field that retention workflows can reference. A churn score in a data warehouse is a research artifact. A churn score in Braze or Salesforce Marketing Cloud is a retention tool.

Acting on predictions without burning budget

The prediction tells you who's at risk. The intervention determines whether you keep them. This is where false positive tradeoffs become real.

Every churn model has a threshold problem. Set the threshold too low (flagging customers at 30% churn probability) and you'll intervene on people who were never going to leave. You'll burn budget on unnecessary discounts and train customers to expect them. Set it too high (only flagging at 80% probability) and you'll catch people too late. By 80% probability, most customers have already made the decision.

The right threshold depends on your intervention cost and customer value. For a high-CLV customer segment, a lower threshold makes economic sense because the cost of losing them justifies more false positives. For a low-CLV segment, a higher threshold avoids wasting retention budget on customers whose future value doesn't cover the intervention cost.

Run the math. Predicted CLV of the at-risk segment times the probability of saving them minus the cost of the intervention. If that's positive, intervene. If it's not, let them go. Not every customer is worth saving, and pretending otherwise is how retention budgets disappear.

The intervention stack

What you do with the prediction matters as much as the prediction itself.

Tier 1: Low risk, high value. These customers aren't likely to churn but they're worth protecting. The intervention is subtle: ensure they're in your best content streams, they get early access to new products, their experience stays frictionless. Don't offer a discount to someone who isn't leaving.

Tier 2: Medium risk, high value. This is where most of your retention budget should go. Personalized outreach, loyalty recognition, proactive problem-solving. At Stanley Black & Decker, the B2B equivalent was a proactive check-in from the account team when engagement signals dipped. Not an automated email. A human touchpoint.

Tier 3: High risk, high value. Aggressive intervention is justified. Retention offers, direct outreach, escalation to customer success. But verify the signal first. High churn probability plus high predicted value is either your most important save opportunity or your model's most expensive false positive.

Tier 4: High risk, low value. Let them go. Seriously. A customer predicted to churn who has low future value doesn't warrant a retention spend. Suppress them from retention campaigns and reallocate that budget to Tier 2.

Maintaining the model over time

Churn models decay. Customer behavior changes, product mix shifts, competitive dynamics evolve. A model that performed well at launch will drift within six months if left alone.

Track two metrics continuously: the AUC (area under the curve) as a measure of discrimination, and the calibration curve to verify that a 70% churn probability actually corresponds to roughly 70% of those customers churning. When either metric degrades beyond your tolerance, retrain.

Set up a monitoring dashboard before you deploy. Not after you notice the model isn't working.

The takeaway

Churn prediction is an operations problem dressed up as a data science problem. The model is the straightforward part. The hard part is defining churn correctly, engineering features that reflect real customer behavior, deploying scores into systems where actions happen, and calibrating interventions so you're spending retention budget where it changes outcomes. Start with one segment, one threshold, one intervention. Measure whether the intervention actually changes churn rates. Then expand.

Keep Reading

Glossary: Churn Rate | Customer Lifetime Value (CLV)