What is propensity scoring?

Propensity scoring is a supervised model that predicts how likely a person is to take a specific action, such as making a purchase, clicking an offer, or churning. It turns a profile and a history of behavior into a single probability, so a system can rank people by how likely they are to do the thing you care about and act on the ones at the top.

How it works

A propensity model is trained on past examples where the outcome is already known: people who did and did not buy, renew, or churn. It learns the patterns in their attributes and behavior that separate the two groups, then scores new people on the same features. The output is a probability for each person, and the usual move is to sort by that score and treat the highest first, whether that means a sales follow-up, a retention offer, or a place in an audience.

Because it only needs labeled outcomes, propensity scoring is cheap to build and widely deployed. It underlies lead scoring, churn flags, and a lot of the “most likely to convert” audiences that marketing and sales tools ship by default.

What it is good for

Propensity is the right tool when the job is prioritization and the cost of acting is low. If you have a fixed number of sales calls to make, calling the highest-propensity leads first is a reasonable use of a scarce resource. If a flag helps a success team notice an account drifting toward churn, the score has done useful work. In these cases you are using the prediction to allocate attention, and being roughly right about who is likely to act is enough.

The trap: prediction is not causation

The mistake that sinks many personalization programs is treating a propensity score as a reason to target someone. A high score says a person is likely to act. It does not say your message, offer, or discount will cause them to act. Target the people most likely to buy and a large share of them would have bought anyway, so the treatment changes nothing for them while the budget is spent all the same. This is sometimes called the “sure things” problem: the highest-propensity people are often the ones who needed no persuasion.

Propensity answers “what will happen.” The question a personalization decision actually needs answered is “what should I do about it,” which is a question about cause and effect. The two come apart most sharply with discounts and retention offers, where spending on someone who would have stayed or bought is pure cost, and where a badly aimed offer can even train people to wait for the next one.

Propensity versus uplift

The causal counterpart to propensity is uplift modeling, which estimates how much a treatment changes the outcome for each person rather than how likely the outcome is. Uplift needs randomized training data and targets the persuadables, the people whose behavior the treatment actually moves, while leaving the sure things alone. If you only have a propensity model, the honest way to check whether your targeting causes anything is to hold out a random share of the audience and compare, which is covered in how personalization is measured.

When propensity is enough

None of this means propensity scoring is wrong. When the action is cheap, the goal is to rank a limited set of people, and you are not claiming the treatment caused the result, a propensity model is a sensible and well-understood tool. The error is only in reading a prediction as a causal promise. Know which question you are answering, and propensity earns its place among the techniques a real system combines.