Configure a Propensity Scoring Model

Propensity scoring transforms historical behavioral signals into per-profile likelihood scores that predict whether a customer will perform a specific action — such as making a purchase, churning, or upgrading — within a defined future window. The core output is a numerical score attached to each profile, together with influential-factor metadata that explains the model's reasoning. Practitioners must decide the prediction goal (conversion vs. churn), the target event field, the look-back window depth, and a re-scoring schedule that balances freshness against compute cost.

Key decisions include: choosing a target variable that reliably signals the desired outcome (e.g., a commerce event rather than a page-view proxy), ensuring the training dataset covers at least two full quarters of seasonality-representative data, and enabling score writeback to the profile store so downstream segments and activation pipelines can consume scores without extra ETL. Model explainability surfaces the top influential factors per cohort, which helps analysts validate that the model is not relying on spurious correlations.

Parallel viability note: This task has moderate parallelism outside AEP. Platforms such as Snowflake Cortex ML and Databricks ML Runtime offer feature engineering and classification model training on behavioral event tables with similar look-back control, and scores can be reverse-ETL'd back to a CDP profile store. Hightouch and most pure-play reverse-ETL tools do not include an embedded ML layer and would require an external model inference service. Teams evaluating non-AEP stacks should budget for a separate MLOps capability and a mechanism to attach scores to unified profiles.