Case Study
Predictive Analytics
Propensity Modelling
Which customers are most likely to take a relevant future action?
Download PDF
A complementary visual resource for a deeper view of the case.
The business challenge
Which customers should be prioritised for acquisition, sales or engagement under limited commercial capacity?
A ranking can improve effort allocation, but optimising response does not always optimise value or incrementality. The objective must match the real decision.
Key business questions
- Which future action should be anticipated, and over what horizon?
- Which population is eligible, and what capacity exists?
- How should probability, value and incrementality be separated?
Data & evidence
- Behaviour before the target event.
- Customer, product, channel and context attributes.
- Eligibility and campaign exposure.
- Value, margin and action cost.
How I would approach it
- Define event, population, horizon and scoring.
- Validate while respecting time.
- Evaluate discrimination and calibration.
- Translate scores into rules with capacity and economics.
What you would get
- Propensity score and prioritisation ranking.
- Opportunity profiles and eligibility rules.
- Thresholds aligned with capacity and economics.
- Validation and monitoring plan.
From analysis to action
Illustrative example: A high score estimates probability under observed patterns; it does not show that contact causes the action.
- Prioritise audiences under explicit constraints.
- Adapt channel, message and frequency.
- Reserve controls and monitor drift.
How success should be measured
- AUC/PR, lift and temporal calibration.
- Incremental conversion and value.
- Cost per action and stability.
Other approaches to consider
Logistic regression offers an interpretable baseline and boosting captures nonlinearity. Estimating who will change because of an intervention requires uplift or experimentation.