Propensity-to-Buy Models

Propensity models use historical data to estimate which customers are more likely to take a specified action, such as buying a product within 30 days. Features may include prior transactions, recency, product views, campaign engagement, and account tenure. The model’s meaning depends on the target definition and observation period. A score of 0.7 is a probability estimate under the model and data, not a promise that a person will buy. High propensity can also mean the customer would purchase without marketing. That differs from uplift, which estimates the incremental effect of an intervention. Targeting solely by purchase likelihood can waste budget on customers who were already going to convert and may over-contact certain groups. Teams should define the action window, label rules, and permitted data, then evaluate calibration and lift on holdout periods. Randomized tests can estimate whether outreach changes behavior. Leakage occurs when features include information created after the prediction time, such as the purchase itself. Seasonal drift, campaign changes, and product availability can affect performance. A score should be one input to a campaign policy that includes contact frequency, consent, exclusions, and a no-contact option. Monitoring should track conversion quality, unsubscribes, complaints, and outcomes across relevant groups. Propensity ranking helps prioritize, but does not establish causation, customer need, or permission to contact. Teams should also document who may receive a score and how suppression rules apply.