Churn Prediction

Churn prediction is the first step toward understanding how machine learning turns historical data into a tool that changes real-world decisions. It introduces the entire pipeline you will use for every supervised problem: define the target, choose features (the raw numbers that describe each customer), train a function that maps those numbers to a probability, evaluate how well it generalises, and act on the result. What you master here — the core meaning of churn, the gap between prediction and action, and the trap of leakage — will carry directly into the neighbouring topics you have noted: Logistic Regression (the model most often used for churn), Rare Event Prediction (where churn's imbalance becomes the whole story), and Decision Trees (an interpretable alternative). Understanding churn prediction now means that when we later build full models, you will already know what the target is and why getting it wrong sank the entire effort.

Think of a streaming service. Every month, some subscribers cancel. The company would love to know, before the month ends, which subscribers are about to leave, so they can offer a discount, a new feature, or a phone call to keep them. Churn prediction is the machine-learning answer to that question. We collect a history of every past subscriber: how long they have been with us, how often they watch, how many devices they use, whether their payment has ever failed. We also know whether each of those subscribers eventually left. From that history, we want to build a rule that, for any current subscriber, reads their numbers and produces a score — say, a number between 0 and 1 — where 1 means 'almost certainly leaving' and 0 means 'almost certainly staying'. The trick is that we do not write the rule by hand; we let the computer discover it from the data. The computer looks at thousands of past examples, finds patterns (e.g., 'subscribers who log in less than once a week are more likely to leave'), and encodes those patterns into a function. Later, we feed new customers through that function and get a churn score. The score itself is not the end goal — the goal is to decide who deserves retention effort. The 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 a set of measurements — heart rate, blood pressure, blood sugar — 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 is exactly the same, but 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 model learns which feature combinations correlate with churn. Once trained, it evaluates a new customer and gives a risk score. The intervention — the treatment — is the retention action. The analogy breaks, though, in a crucial way. A doctor's model is built on centuries of medical knowledge and causal reasoning: we know why high blood pressure causes damage. A churn model is purely empirical: it finds correlations without understanding why. It might learn that customers whose names start with 'A' churn more, which is almost certainly a meaningless artefact of the training data. Worse, the model reflects the past; if the company changes its product or pricing, the old patterns may stop holding. The doctor's model is also more robust because the underlying physiology does not change overnight, whereas customer behaviour can shift with a single competitor's promotion.

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.

This concept sits at the very start of your Machine Learning Foundations study. You have noted in your library that the overarching goal is learning functions from data rather than specifying them by hand — churn prediction is your first concrete instance of that. The supervised/unsupervised/reinforcement split comes to life here: churn is a supervised problem because we have labelled examples (churned or not). The train-validate-test discipline, which you will soon study in depth, is the only way to trust a churn 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, then a model that always predicts 'stays' is 98% accurate but completely useless. That imbalance is a theme you will revisit throughout your notes. Finally, the classical models you have already recorded — Logistic Regression, Decision Trees, and Random Forests — are the standard tools for churn on tabular data; your library's emphasis that *classical models remain the right answer for small tabular data* is precisely the context here.