Churn Propensity: How to Predict Customer Churn
- Churn propensity is the probability that a customer stops buying from you, estimated from behavior such as recency, frequency, spend and complaints.
- In most online stores customers never cancel, so you must define churn as an inactivity window that fits your purchase cycle before you can predict it.
- Logistic regression, decision trees, random forests, gradient boosting and survival analysis all predict churn; survival analysis also predicts when it will happen.
- A churn model outputs a probability between 0 and 1; the threshold that turns it into high or low risk is a business decision about retention budget.
- A churn model shows who is likely to leave, not who will respond to an offer, so test retention tactics on random sub-groups with a control group.
Churn propensity is the likelihood that a customer will stop buying from you. A propensity-to-churn model estimates that likelihood for every current customer: it learns the patterns shared by customers who already left, then scores active customers by how closely they match those patterns. You get a ranked list of who is at risk, early enough to act.
Riddle me this: what is the common link between highly successful organizations (no matter the industry) and highly successful individuals (no matter their field)? They all have an extremely clear vision of their journeys: where they were, where they are, and where they're headed.
In eCommerce and retail, that vision comes from your customer data. Your store already owns an abundance of zero- and first-party data, and one of the most useful things you can do with it is predict customer churn.
This guide shows what churn propensity is, how a churn prediction model works, which modeling methods to use, how to build and score one step by step, and how to use the scores to keep your high-value customers. For the causes of churn and the churn rate formula, see our guide on how to reduce customer churn.
What is churn propensity?
A churn propensity estimate takes into account what customers do, such as their payment history or buying frequency, and what they tell you, to predict whether a customer is likely to leave.
Other factors that influence a customer's propensity to churn:
- Customer satisfaction with products, delivery and returns.
- The effectiveness of customer service when something goes wrong.
- The competition, including better offers and lower prices elsewhere.
eCommerce businesses (but not only them) use predictive analytics, machine learning and data science to predict and prevent churn, grow their retention rate, and grow their business in a healthy way. Churn propensity models are built on the factors most likely to cause churn and on a company's ability to address them.
What is a propensity-to-churn model?
In the beginning, there was the acquisition stage. Organizations would bend over backward to acquire new customers but didn't care much about retaining them. There were plenty of fish in the sea, so to speak.
Today, the situation is as different as night and day. When acquisition stops being sustainable, smart companies become more defensive and focus on retention.
This is where the churn model comes in handy, acting as a "crystal ball" for churn risk inside your customer base:
- It analyzes your historical data, investigating customers who have already stopped purchasing your products.
- It searches for patterns or shared attributes in these customers.
- Based on those patterns, it predicts which current customers have the highest propensity to leave.
A churn model is a prediction, not a certainty. It tells you who looks like the customers you lost, not why each individual customer will leave. Pair the scores with churn analysis and customer feedback to understand the reasons.
How do you define churn when customers never cancel?
A model can only predict what you tell it to predict. The churn definition decides which past customers count as "churned" in your training data, so it shapes every score the model produces.
- Contractual (subscription) businesses: churn is observed. A canceled plan or an unrenewed subscription is the label.
- Non-contractual (most online stores): churn is inferred. Pick an inactivity window that fits your purchase cycle. A store where customers reorder every month might use 90 days; a furniture store might need a year or more.
Check the window against your own data: look at how long it normally takes a returning customer to place a second or third order, and set the threshold well beyond that. A window that is too short labels loyal but slow buyers as churned. RFM segmentation is a practical starting point, because recency (days since the last order) is the most direct signal of silent churn.
What does a propensity-to-churn model look like in practice?
Suppose you operate in the Health & Wellness industry, selling proteins and other supplements through a monthly subscription. (The logic of this example is the same in any industry.)
You want a propensity-to-churn model that predicts which customers will cancel their monthly subscriptions. Your first task is gathering historical data on customers who previously unsubscribed. Look for information such as:
- How long the customer was subscribed
- How much the customer spent each month
- The customer's age and gender
- The type of subscription the customer had
- Any complaints from the former customer
After gathering this data, you feed it to a churn scoring algorithm, which predicts the probability of each current customer canceling their subscription. The results might look like this:
- Low risk: a 25-year-old customer who has been with the company for three years. The monthly order is around $40, and there have been no complaints. This pattern suggests the customer is unlikely to churn.
- High risk: a 55-year-old customer who has been with the company for half a year, spends around $12 per month and complains regularly. The model might predict a high propensity to leave in the near future.
In practice, behavior (tenure, spend, complaints) usually carries the signal. Use demographic fields with care: they can help a model, but you should not treat a customer as a churn risk only because of their age or gender. Once you find the high-risk customer segment, you can act on it.
Tl;dr: to see which customers are most likely to stop coming to your store, first look at the data on customers who have already left. Find the patterns and similarities, then identify the current customers who share the most characteristics with those who left.
Which methods predict churn propensity?
No method is best for every store. The right choice depends on how much data you have, whether you need to explain the result to your team, and whether you care about if or when a customer churns. Most teams start with a simple, explainable model and move to a more complex one only if it predicts clearly better on test data.
| Method | What it does | Best for | How to read it |
|---|---|---|---|
| RFM scoring | Scores customers on recency, frequency and monetary value with rules, not a trained model | A fast baseline, or stores without data science resources | Falling recency in a previously frequent, high-value customer is the warning sign |
| Logistic regression | Estimates a churn probability from a weighted combination of customer features | A first model you need to explain to stakeholders | Each coefficient shows whether a feature raises or lowers churn risk |
| Decision tree | Splits customers into groups with simple if-then rules | Readable rules the team can act on directly | Follow a branch to see which conditions define a high-risk group |
| Random forest / gradient boosting | Combines many trees to capture non-linear patterns and interactions | Larger datasets where predictive power matters most | Use feature-importance outputs to see which signals drive risk |
| Survival analysis (for example, Cox models) | Models the time until churn, including customers who have not churned yet | Planning when to intervene, not only whom to target | Read the hazard: how risk rises or falls over the customer's lifetime |
Academic research explores these approaches in depth. For example, Braun and Schweidel's 2011 paper Modeling Customer Lifetimes with Multiple Causes of Churn (MIT DSpace) models customers who leave for different reasons. One practical takeaway: removing one cause of churn may add less customer lifetime than you expect, because customers can still leave for other reasons.
How do you create a propensity-to-churn model?
Propensity-to-churn models are used to improve customer retention and to protect customers with high lifetime value. In any field, prevention is cheaper than treatment, so it's worth knowing how to build such a model: it empowers you to act before it's too late.
-
Collect historical dataGather data on your customer base, including both customers who left and active customers. Include demographic info, purchase history and buying patterns, and label each past customer as churned or not using your churn definition.
-
Create a training set and a test setSplit your data in two. The training set teaches the model to predict churn. The test set, which the model never sees during training, measures how well the model performs.
-
Pick a machine learning modelChoose a method such as logistic regression, a decision tree or a random forest (see the table above). Pick the one that fits your data and your need to explain results, then train it on your training set.
-
Test the modelUse the trained model to score the customers in your test set. Each customer gets a churn probability between 0 and 1, which you can convert into a churn or no-churn prediction with a threshold.
-
Evaluate the modelCompare the predictions with what actually happened. Look beyond accuracy: churners are often a minority, so also check precision (how many flagged customers really churned) and recall (how many churners you caught). Run multiple tests until you're satisfied that the model can predict customer churn.
-
Run the model on your customer baseWhen training and testing are complete, score your current customers and identify the ones at risk of churning.
-
Adjust the model regularlyCustomers come and go, so you'll never be short of new data. Keep collecting it, retrain the model, test alternative methods and fix inaccuracies. Customer behavior changes over time, and a model that is never retrained slowly loses accuracy.
Treat your propensity-to-churn model like a Tamagotchi: it can run on its own, but remember to check in on it from time to time.
How do you calculate propensity to churn?
An important disclaimer before we go into propensity scoring: any propensity-to-churn model is based on an algorithm. You'll need both qualitative and quantitative data to train it, and the model can only predict propensity well if you pay attention to the training phase and feed it enough data to reach relevant conclusions.
Churn prediction is a binary classification problem: the outcome is 1 (churns) or 0 (stays). Most models don't output the 0 or 1 directly. They output a probability, and you decide where to draw the line. Logistic regression, for example, uses this formula:
Churn probability = 1 / (1 + e−z), where z = b0 + b1x1 + b2x2 + …
Here, x1, x2… are customer features and b0, b1… are the weights the model learns from your historical data. The table below uses made-up weights to show how the arithmetic works.
| Customer | Days since last order | Orders in last 12 months | Complaints | z | Churn probability |
|---|---|---|---|---|---|
| Customer A | 120 | 2 | 1 | −3.0 + 2.4 − 0.8 + 0.8 = −0.6 | 1 / (1 + e0.6) ≈ 0.35 (35%) |
| Customer B | 20 | 6 | 0 | −3.0 + 0.4 − 2.4 + 0 = −5.0 | 1 / (1 + e5.0) ≈ 0.007 (under 1%) |
Afterward, each customer goes into a segment:
- High churn risk: probability above your threshold (with a 0.30 threshold, Customer A is high risk).
- Low churn risk: probability below it (Customer B).
Where you put the threshold is a business decision. A low threshold catches more future churners but also flags more customers who would have stayed anyway, which costs you retention budget. Many teams use more than two tiers so they can match the cost of an intervention to the level of risk.
How do you reduce customers' propensity to churn?
Churn modeling gives you an ace in the hole: a way to monitor customer behavior and pinpoint the signals that come before a customer dumps your brand, then recognize those signals in current customers. If you can clearly see the signs of departure, you can quickly put prevention tactics in motion and minimize customer churn.
Another essential disclaimer for this section: your churn model doesn't identify the customers most likely to be persuaded to stay. Even if it reveals customers at risk of leaving, it won't show you who is most receptive to your retention strategy. Some high-risk customers will leave whatever you do; others would have stayed without an offer. Predicting who responds to an intervention is a different technique, usually called uplift modeling.
You can, however, use your propensity-to-churn scores to find out whether your churn reduction strategies work:
-
Pick a high-risk segmentStart with customers who have a high churn propensity and a high lifetime value. They are the most costly to lose.
-
Split it into sub-groups at randomGive each sub-group a different retention tactic, and keep one sub-group as a control group that receives nothing new.
-
Compare the outcomesAfter a set period, compare repeat purchases and churn across the groups. Because the groups started with the same risk, you compare apples to apples and don't get sidetracked by subjective interpretations.
A/B testing can work wonders here, and it doesn't need to be limited to website personalization or email subject lines. Churn prediction models are incredibly helpful in experimentation: before you set a retention strategy in stone, you can use sub-groups to control and measure its effect.
Strip away the buzzwords, and lowering the propensity to churn comes down to watching the little (but crucial) things in customer experience. Here are a few ways to prevent and reduce customer churn in eCommerce, retail, or any other business:
Make customer service a priority
No matter how high-tech you get with your churn analysis and prevention, nothing beats fast and efficient problem-solving. Never underestimate customers' need to feel seen, heard and valued. Invest in training your customer service representatives, provide live chat support, and prioritize tickets from high-value customers.
Learn from complaints
Customer feedback is your mirror. It shows your strong and weak points without vanity metrics, ego-boosting or pampering. Act on customer complaints and bad reviews, and fix repeat problems for good. This way, you don't frustrate customers with the same problem over and over, and you prove you're invested in better customer experiences.
Turn your customer base into a community
Humans have evolved to live in communities, and we still crave a sense of belonging to something bigger than ourselves. Use your communication channels to build a community around your brand. Give customers an identity they're comfortable with, and turn that sense of belonging into a reason to stay.
Provide personalized experiences
Data doesn't stop at churn modeling. Use it to provide tailored experiences that feel relevant and valuable. For example, analyze buyer behavior, look for patterns in buying frequency or product assortment, and send personalized product recommendations.
Churn reduction isn't a one-time project that you tackle today and forget next quarter. It's a continuous effort to keep your customers happy and engaged. For a full retention plan, see our customer retention strategy guide, and for how retention connects to long-term value, read how to grow Customer Lifetime Value.
How can you detect churn risk without building a model?
A trained propensity model is powerful, but it needs clean data, time and skills. RFM segmentation gives you much of the practical value with far less effort. A customer who used to buy often and spend a lot, but whose recency is slipping, is exactly the customer a churn model would flag.
Nexus by Omniconvert puts that approach to work:
- RFM segmentation of your customer base, updated from your store data (see the RFM segmentation guide).
- CLV tracking by segment, cohort and acquisition channel, so you can see which sources bring customers who churn after one order.
- Churn-trigger products: Nexus flags first-order products whose buyers rarely come back, so you can remove them from acquisition campaigns.
- Direct activation: push at-risk segments to Meta Ads, Google Ads and Klaviyo for targeted win-back campaigns.
Use RFM segments as your baseline, and add a trained propensity model when you have the data and the need for finer predictions. The two work well together: the model ranks individual risk, and the segments tell you which group each customer belongs to and what they are worth.
See which valuable customers are drifting away, with RFM segmentation and CLV tracking.
See Nexus by Omniconvert →Frequently Asked Questions
Churn propensity is the probability that a customer will stop buying from your brand within a defined period. It is estimated from behavior such as purchase recency, buying frequency, spend and complaints. A high churn propensity marks a customer you are likely to lose unless you act.
Churn is the act of a customer leaving you. Churn analytics measures the rate at which customers leave your brand and uncovers the reasons behind it, so you can predict and prevent future churn.
Define what churn means for your business, gather historical data on former and active customers, and split it into a training set and a test set. Train a method such as logistic regression or a random forest, evaluate it on the test set, then score your current customers and retrain the model regularly.
A churn model outputs a probability between 0 and 1 for each customer. Churn prediction is a binary classification problem, so you set a threshold that turns the probability into a churn or no-churn label, or you group customers into risk tiers such as high and low churn risk.
Propensity modeling uses statistics or machine learning to predict how likely a customer is to take a specific action, based on historical data. Churn propensity models predict the likelihood of leaving; other propensity models predict actions such as purchasing or responding to an offer.
No single algorithm is best for every business. Logistic regression is simple and easy to explain, tree ensembles such as random forest and gradient boosting capture more complex patterns, and survival analysis predicts when a customer is likely to churn. Start simple and switch only if a more complex model predicts clearly better on your test data.
In a non-contractual business, customers don't cancel; they stop buying. Define churn as no purchase within an inactivity window that fits your purchase cycle, set well beyond the normal time between orders. RFM segmentation helps, because declining recency is the most direct sign of silent churn.
Not on its own. A churn model shows who is likely to leave, not who will respond to a retention offer. Test your tactics on random sub-groups of an at-risk segment with a control group, or use uplift modeling, to find the customers your intervention actually persuades to stay.
It used to be difficult to predict when a customer would leave. Machine learning helps you spot patterns and red flags, and predict customer behavior before it shows up in your revenue. Define churn for your business, start with a simple, explainable model or with RFM segments, and use the scores to test retention tactics against a control group. Combine the insights of a well-trained algorithm with the personal touch of human interactions, and you get a retention strategy that is hard to beat. Good luck, and remember to have fun!
Spot churn risk before customers leave
Nexus by Omniconvert automates RFM segmentation, cohort analysis and CLV tracking, so you can see which valuable customers are drifting away. Push at-risk segments directly to Meta Ads, Google Ads and Klaviyo for win-back campaigns.