Churn Prediction

Churn prediction teaches you how to use past history to change real-world decisions before it is too late. Imagine a subscription service losing customers every month. If the company knows who will leave next, they can offer a discount or a call to keep them, saving revenue. To do this safely, you must avoid 'leakage' (a critical error where information from the future unknowingly slips into the past data, making a model appear perfect when it is actually broken). This process also illustrates 'supervised' learning (the method of teaching a computer by showing it labeled examples, like 'this customer left' or 'this customer stayed'). Mastering this gives you a concrete framework for evaluating whether a prediction will actually work in the real world, which applies to every Logistic Regression model (a specific statistical method for binary choices) and Decision Trees (a flowchart-like structure that splits data based on simple rules) later on.

Think of a streaming service like Netflix or Spotify. Every month, some subscribers cancel their account. The company wants to know, before the month ends, which specific people are about to leave, so they can intervene with a special offer or a friendly phone call. To build this ability, the company collects history for every past subscriber: how long they have been with us, how often they watched, how many devices they use, and whether their payment has ever failed. Crucially, the company also knows the final outcome for each of those people: did they eventually leave or stay? From this history, we want to build a rule that, for any current subscriber, reads their numbers and produces a score between 0 and 1. A score of 1 means 'almost certainly leaving' and 0 means 'almost certainly staying'. We do not write this rule by hand. Instead, we let the computer discover it from thousands of past examples. The computer finds patterns, such as 'subscribers who log in less than once a week are more likely to leave,' and encodes those patterns into a function. When we feed new customers through that function, we get a churn score. This prediction is useful only if we can intervene in time, with a cost less than the revenue the customer would bring if they stayed.

A doctor does not wait until a patient collapses to treat them. Instead, the doctor looks at measurements like heart rate and blood pressure and compares them with years of records from thousands of other patients. From those records, the doctor has learned a pattern: certain combinations of measurements are associated with a higher risk of a heart attack. When a new patient comes in with slightly abnormal numbers, the doctor flags them as high-risk and prescribes preventive medication. Churn prediction works exactly the same way: the patient is a customer, the heart attack is a cancellation, and the doctor is a machine-learning model. The historical records are the training data: every past customer, their features, and their eventual outcome. The analogy breaks in one crucial way. A doctor's model is built on medical knowledge of cause and effect; a churn model is purely empirical, finding correlations without understanding why. It might mistakenly learn that customers whose names start with 'A' churn more, which is meaningless noise rather than a real cause. This caveat warns us that correlation does not always mean we can change the outcome.

Churn prediction is the process of building a function from historical customer data that estimates, for each current customer, the probability that they will stop using the product within a defined time window, so that retention actions can be targeted at those most likely to leave. It is a supervised learning task because it relies on past examples where the outcome was already known.

This concept sits at the very start of your Machine Learning Foundations study as your first concrete instance of learning functions from data rather than specifying them by hand. Churn prediction is a 'supervised' problem (a category where we teach a model by showing it labeled examples, such as 'churned' or 'not churned'), which is why the train-validate-test split (a method of dividing data into three parts: one for learning, one for tuning, and one for final testing) is essential to trust the model. The neighbouring topic of Rare Event Prediction is directly relevant because churn is often a rare event; if only 2% of subscribers leave each month, a simple model that always predicts 'stays' would be 98% accurate but completely useless for identifying who actually leaves. Finally, the classical models you have recorded — Logistic Regression, Decision Trees, and Random Forests — are the standard tools for churn on tabular data, reinforcing the principle that simple models often remain the best answer for structured customer information.