Propensity modeling estimates how likely a person or account is to take a specific action. In audience targeting, these scores help decide who to reach, what to say, and when to act.

Core building blocks

• Defined outcome: One clear behavior such as purchase, upgrade, sign-up, churn, or engagement.
• Training data: Historical observations that include the outcome and relevant predictors.
• Modeling method: Techniques such as logistic regression, gradient boosting, random forests, or calibrated neural nets.
• Propensity score: A 0–1 probability calibrated to reflect real-world likelihoods.

From score to action

• Segment by score: Create tiers such as high (0.7–1.0), medium (0.4–0.69), and low (0–0.39).
• Tailor experience: Different offers, creative, and sequences by tier.
• Optimize spend: Bid aggressiveness and channel mix aligned to expected value.
• Triggering: Real-time or near-real-time actions when scores or signals change.

What "good" looks like

• Well-calibrated: If 100 people have a 0.3 score, about 30 should do the action.
• Stable features: Predictors that are available, compliant, and not easily gamed.
• Explainable: Clear drivers help creative and sales teams act with confidence.

Launch with a practical, testable plan that moves from data to impact.

Step 1: Frame the use case

• Define the decision: What will change if a score is available?
• Outcome window: Choose a prediction horizon (for example, 30, 60, or 90 days).
• Eligible population: Clarify who is scored and who is out of scope.

Step 2: Data and features

• Signals: Recency, frequency, monetary value, product usage, lifecycle stage, channel engagement, firmographics, and contextual signals.
• Leakage control: Exclude features that directly encode the future outcome.
• Freshness: Decide update cadence (daily, weekly, or event-driven).

Step 3: Modeling and validation

• Baselines first: Start with logistic regression and compare to tree-based models.
• Holdouts: Keep a time-based holdout set for realistic evaluation.
• Metrics: ROC-AUC for ranking, calibration curves, lift and gain charts, and decision-oriented metrics like expected profit.

Step 4: Operationalization

• Scoring pipeline: Automated jobs that produce scores and confidence intervals.
• Activation: Sync to ad platforms, CRM, MAP, and onsite personalization.
• Experimentation: A/B test score thresholds, offers, and channels before wide rollout.

Step 5: Communicate the playbook

• Score tiers: Definitions, recommended treatments, and suppression rules.
• Creative guidance: Messaging frameworks aligned to predicted intent.
• Escalation paths: What to do when scores drop or spike.

A durable program protects privacy, measures impact, and keeps improving.

Responsible data and compliance

• Consent-aware: Only use data collected with clear permissions.
• Fairness checks: Monitor for disparate performance across segments.
• Feature governance: Track lineage of every field used in the model.

Measuring business value

• Incrementality: Use randomized controls or geo-experiments to isolate lift.
• Threshold economics: Choose cutoffs using expected value and capacity constraints.
• Attribution sanity checks: Compare model-driven outcomes to holdout benchmarks.

Lifecycle management

• Retraining cadence: Refresh when input drift or performance decay appears.
• Monitoring: Track calibration, lift by decile, and data freshness alerts.
• Model portfolio: Maintain distinct models for different outcomes and stages.