Adaptive Models
This domain tests how Pega's self-learning models produce propensity, how they learn from outcomes, and how you judge whether they are performing.
How adaptive models work
- A model is created per action, per channel and per direction; you do not train them offline first.
- They start neutral (roughly a 50% propensity) and update after every recorded outcome - impression plus accept or reject - from interaction history.
- Predictors come from customer properties; symbolic predictors are grouped into bins, numeric ones into ranges.
- Every model that participated in a decision receives the outcome feedback, not only the action that was displayed.
- Performance is monitored with metrics such as AUC and response rate, plus predictor-level transparency showing which factors push the score up or down.
Worked example
A new mobile-data upsell action goes live with no history, so its model outputs about 0.50. After a week of impressions, the accept rate is 8% overall but much higher for customers with high data usage and short tenure. The model shifts its propensity upward for those predictor combinations, and a later strategy run ranks the upsell higher for that segment. Reasoning: the model did not need a data-science cycle; the outcome records themselves moved the score.
Traps that cost marks
- Saying adaptive models need historical training data before they can score.
- Believing only the winning action's model learns.
- Confusing AUC (ranking quality) with response rate (share of accepts).
- Expecting models to learn during simulation testing; simulation does not update models.
- Ignoring that a model is scoped to one action, channel and direction.