Propensity model

A propensity model assigns each customer or record a score representing the estimated probability that they will take a specific action, purchasing, churning, responding to an offer, defaulting on a loan, based on patterns learned from historical examples of people who did and did not take that action. It is typically built as a classification model, often logistic regression or a tree-based method, trained on past behavior and outcomes.

Propensity scores differ from uplift modeling in an important way: a propensity model predicts the likelihood of an outcome regardless of any intervention, while uplift modeling predicts the incremental effect of an intervention, such as a discount, on that likelihood. A high propensity to buy does not mean an offer caused the purchase; some high-propensity customers would have bought anyway.

Propensity models matter because they let teams prioritize limited resources, marketing spend, retention outreach, collections effort, toward the people most likely to respond, and they feed directly into next-best-action and recommendation system logic. Common pitfalls include using propensity scores to target people who would have converted regardless, wasting the intervention on already-likely outcomes, and letting the model drift out of date as customer behavior and the underlying churn rate change.