Churn Prediction Models: How Subscription Brands Spot Who Will Cancel

A churn prediction model tells you which subscribers are likely to cancel before they do. Most guides on the topic are written for B2B software teams with data scientists and account managers. This one is for subscription ecommerce brands: which signals you already have, when rules beat machine learning, how to test a model honestly, and what you're allowed to do with the score.

What is a churn prediction model?

A churn prediction model gives each active subscriber a score: the likelihood that they churn within a window you choose. It learns from past subscribers whose outcome you already know, the ones who cancelled and the ones who stayed, and looks for the patterns that came before each outcome.3Braze Docs, Predictive Churn (read October 14, 2026)

The output is a number per customer. Klaviyo exports churn risk as a value from 0 to 1 and describes it as "the probability of a customer churning", based on "their number and frequency of orders".1Klaviyo Help Center, Guide to Klaviyo's predictive analytics (read October 14, 2026) Braze assigns a "churn risk score between 0 and 100".3Braze Docs, Predictive Churn (read October 14, 2026) Whatever the scale, the score answers one question: who should we look at first?

A model doesn't tell you why a customer will leave, and it doesn't keep anyone. It sorts customers. What you do with that order, and whether it pays, is a separate question we cover further down. If you only need to know how many subscribers you lost last month, that's churn rate calculation, not prediction.

Define churn before you predict it

Every model learns one definition of churn. If the definition is vague, the model learns noise. Three choices come first:

1. The event. A cancellation, a renewal that failed and was never recovered, or a customer who stops ordering. In subscription ecommerce, pauses complicate this: a three-month pause is not a cancellation, but a pause that's extended twice may be one in all but name.
2. The window. "Will cancel in the next 60 days" and "will cancel before the next renewal" are different models. Braze lets you define churn as an action a user does or stops doing, over a window of up to 60 days.4Braze Docs, Create a churn prediction (read October 14, 2026)
3. The population. Active subscribers only, or also one-time buyers? First-cycle subscribers churn for different reasons than two-year subscribers, and often deserve their own model.

History matters more than model choice at this stage. Pendo's guide for its churn model asks for historical data "two to three times the prediction window", and enough history to capture 200 churned and 500 renewed accounts.5Pendo Help Center, Build a churn model in Pendo Predict (read October 14, 2026) Those are B2B account numbers, but the logic holds: you need enough past churners to learn from.

Keep voluntary and involuntary churn apart

A customer who clicks cancel and a renewal declined for insufficient funds both end a subscription. They are not the same problem, and one label for both teaches the model a blurred target. Stripe stores the reason a customer gives when cancelling through its portal, which gives you a clean label for voluntary churn.13Stripe Docs, Add a cancellation page to the customer portal Failed renewals have a decline code instead. Label them separately, model them separately if you model both, and send them to different fixes. We cover the payment side in subscription payment failures.

Churn signals that work for consumer subscriptions

B2B churn guides talk about logins, seats and feature adoption. A brand selling coffee, skincare or supplements by subscription has different signals, and most of them already sit in the subscription app, the checkout and the helpdesk.

Two rules make these signals useful. First, change over time beats a single value: "skipped 2 of the last 3 orders" says more than "skipped once ever". Second, every signal must be measured as of the scoring date, from data that existed then. That second rule is where most models break, as we'll see in the pitfalls.

Start simple: rules and RFM scores

Before any machine learning, a few rules written with your retention team will catch a large share of at-risk subscribers, and give you a baseline any model must beat.

RFM is the classic version. It scores customers on recency (days since the last order), frequency (how often they order) and monetary value (how much they spend). Klaviyo's RFM report ranks each customer's percentile on each measure, converts it into a score from 1 to 3, and assigns a group such as Champions, Loyal, Needs attention, At risk or Inactive. "At risk" is defined as "a customer that has not purchased recently and/or tend to spend less overall".2Klaviyo Help Center, Understanding scoring and customer groups in the RFM report (read October 14, 2026) Free orders, such as 100% discounted ones, are left out of the analysis.2Klaviyo Help Center, Understanding scoring and customer groups in the RFM report (read October 14, 2026)

RFM was built for repeat purchase. For subscriptions, add the events RFM can't see. A rule set might look like this:

Rules have two strengths over models: anyone can read them, and they work from the first week. Their weakness is that they don't weigh signals against each other. A subscriber with one skip, one ticket and a discount habit may be riskier than one with two skips, and only a model can tell you that from your data.

Machine learning models, explained plainly

When you have enough history and enough past churners, a statistical model can weigh dozens of signals at once. Three families cover almost every subscription use case.

Logistic regression

Logistic regression is the standard starting point because its output is a probability and each input has a visible weight. scikit-learn notes that logistic regression is "more likely to return well calibrated predictions by itself", meaning a score of 0.8 tends to match an 80% churn rate among similar customers, provided the model is well specified and tuned.10scikit-learn, Probability calibration That property matters when you use the score to decide how much an offer is worth.

Gradient boosting

Gradient boosting stacks hundreds of shallow decision trees. It captures patterns a linear model misses, like a discount habit that only matters after the third order. scikit-learn's histogram-based version can be "orders of magnitude faster" than the classic one on tens of thousands of samples or more, and has "built-in support for missing values", which helps when a subscriber has never opened a ticket or never received an SMS.11scikit-learn, Ensembles: histogram-based gradient boosting The trade-off: its scores are not as well calibrated as logistic regression's, so check calibration before you put a price on a score.10scikit-learn, Probability calibration

Survival analysis

Classification asks "will this subscriber churn in the next 60 days?" Survival analysis asks "how long until this subscriber churns?" It was built to measure lifespans, and the lifelines documentation uses users subscribing to a service as an example.12lifelines documentation, Introduction to survival analysis Its key idea is censoring: a subscriber who is still active hasn't churned yet, but may later. The same documentation warns that ignoring these customers makes you "severely underestimating the true average lifespan".12lifelines documentation, Introduction to survival analysis For brands with long-lived subscribers, survival curves by cohort are often more useful than a single churn score.

Inside Tagada

Build the funnel once, route every customer to the right offer

Checkout, upsells and subscriptions on one flow, with conditions per country, order or customer.

Start free

Build or use the score your tools already ship

You may not need to build anything. Several tools a subscription brand already uses ship a churn score, each with its own definition and data requirements.

Sources: vendor pages, read October 14, 2026.1Klaviyo Help Center, Guide to Klaviyo's predictive analytics (read October 14, 2026)3Braze Docs, Predictive Churn (read October 14, 2026)4Braze Docs, Create a churn prediction (read October 14, 2026)5Pendo Help Center, Build a churn model in Pendo Predict (read October 14, 2026)6Subbly, Predictive Churn AI launch announcement (September 13, 2023, updated October 9, 2026) Braze retrains its model "every two weeks automatically" and charges extra for daily score updates.4Braze Docs, Create a churn prediction (read October 14, 2026) Klaviyo doesn't name its method; it says it "applies a combination of data science and machine learning techniques".1Klaviyo Help Center, Guide to Klaviyo's predictive analytics (read October 14, 2026)

Vendors also publish accuracy figures for their models. Subbly's 2023 launch reported accuracy "on training data", which says little about performance on new customers.6Subbly, Predictive Churn AI launch announcement (September 13, 2023, updated October 9, 2026) Treat any vendor accuracy number as measured their way, on their data, and test the score on your own subscribers before you trust it.

The limit of email-platform scores is what they see. Klaviyo's churn risk is built from orders.1Klaviyo Help Center, Guide to Klaviyo's predictive analytics (read October 14, 2026) It doesn't know about skips, pauses, tickets or an expiring card unless you send those events in. A score built on your subscription and payment data will usually see more.

How to evaluate a churn prediction model

A model that looks good in a notebook can be useless in production. Four checks keep you honest.

Don't trust accuracy

If 5% of your subscribers churn in a month, a model that predicts "nobody churns" is 95% accurate and catches no one. Use precision and recall instead. scikit-learn defines precision as "the ability of the classifier not to label as positive a sample that is negative", and recall as "the ability of the classifier to find all the positive samples".8scikit-learn, Metrics and scoring: precision, recall and ROC In churn terms: precision is the share of flagged subscribers who actually churn, recall is the share of churners you flagged.

In this example, precision is 300 ÷ 1,000 = 30%. Recall is 300 ÷ 500 = 60%. Accuracy is 91%, which says almost nothing.

Measure lift in the riskiest group

Lift compares the churn rate in your flagged group with the average. Here, 30% churn among flagged subscribers against 5% overall is a lift of 6: contacting the top 10% reaches six times more future churners than contacting 10% at random. Lift is the number that tells a retention team whether the list is worth working, and it maps directly to budget: you only pay for offers on the group you contact.

Test on a later period

Churn data has a time order. scikit-learn's documentation for TimeSeriesSplit warns that ordinary cross-validation is inappropriate for time-ordered data because it "would lead to training on future data and evaluating on past data".9scikit-learn, TimeSeriesSplit Train on, say, January to June, and test on July to September. If performance drops sharply on the later period, the model learned something that didn't last.

Check calibration if you price offers on the score

If you decide that a $10 offer is worth it above a 0.4 churn probability, 0.4 has to mean roughly 40%. scikit-learn describes a well calibrated classifier as one where, among customers scored near 0.8, "approximately 80% actually belong to the positive class".10scikit-learn, Probability calibration Plot predicted against actual churn by score band before you set thresholds.

How to act on churn scores

A score changes nothing until it triggers an action. The useful pattern combines the score, which says who, with the signals, which say what. A high score driven by skips calls for a frequency change; a high score driven by an expiring card calls for a card update.

The holdout is what separates prediction from retention. A model can predict churn perfectly and the offer can still change nothing, or only delay cancellation by one cycle. Without a group that got no offer, you can't tell the difference. For the tools that run these offers, see churn prevention software.

Save offers within the legal limits

A churn score can tell you who to show an offer to. It can never justify making cancellation harder for high-risk customers. The rules are the same whatever the score says; the full picture is in our guide to negative option billing.

Two more traps are specific to score-driven offers. A discount that ends is a promotional price: Visa requires a notice at least 7 days before the payment when a trial, introductory offer or promotion ends, with the amount, the date and a simple way to cancel.17Visa Core Rules and Visa Product and Service Rules, April 18, 2026 And a customer who restarts after cancelling is a new sign-up: ROSCA requires express informed consent before charging them again.16Restore Online Shoppers' Confidence Act, 15 U.S.C. § 8403

Pitfalls that make churn models lie

Data leakage. scikit-learn defines it as information "that would not be available at prediction time" being used to build the model, which produces "overly optimistic performance estimates".7scikit-learn, Common pitfalls and recommended practices: data leakage In subscriptions, the classic leaks are a feature like "visited cancellation survey" or "subscription status" computed after the scoring date, or a "days since last order" measured at export time instead of at scoring time. The fix: build every feature as of the scoring date, and split your data before any preprocessing.7scikit-learn, Common pitfalls and recommended practices: data leakage

Involuntary churn mixed with voluntary. If failed renewals sit in the same label as cancellations, the model learns that "card expiring soon" predicts churn, and the retention team ends up sending discounts to customers who needed a card update. Stripe's Smart Retries, for example, don't retry hard declines like lost_card or stolen_card until a new payment method is added.14Stripe Docs, Automate payment retries (Smart Retries) No offer fixes that. Route payment risk to dunning and the renewal failure fixes, and keep the voluntary model for voluntary churn.

Rare churners. With a low churn rate, a model can look accurate while catching almost no churners. Compare models on precision, recall and lift, as above.

Censored subscribers. A subscriber who joined last month and is still active hasn't proven loyal. Dropping or mislabeling recent subscribers biases the model; survival analysis handles them by design.12lifelines documentation, Introduction to survival analysis

Your own offers change the data. Once you act on scores, the subscribers you saved look like non-churners in the next training set. Keep the holdout group running, and log every offer so the model can see it.

Drift. A new price, a new product or a new cancel flow changes behavior. Retrain on a schedule, and re-check performance on the latest period each time.

Where Tagada fits

Whatever tool computes the score, the order stays the same: define churn and split it by cause, start with rules, test any model on a later period, act through a holdout, and never let a score stand between a customer and the cancel button.