Decision Management: Models, Strategies, and Arbitration
Decision Management
Decision Management tests how predictive and rule-based logic combine to choose actions. Expect decision strategies, decision tables, decision trees, scorecards, adaptive models, Next-Best-Action, arbitration, and interaction history. The data scientist must build models that are deployable, interpretable, and measurable in production.
Core rules
- Decision strategies orchestrate: filter eligibility, score with models, then arbitrate by value, cost, and constraints.
- Scorecards are transparent and easy to explain; adaptive models update online from feedback.
- Predictive models require consistent features at training and scoring; PMML or Pega-native models must be validated.
- Interaction history supplies outcomes; without feedback, adaptive models cannot learn.
- Arbitration is not highest propensity wins; it balances propensity, value, cost, and business rules.
Worked example
A retention strategy has three offers. The propensity model scores 0.8, 0.6, and 0.5. Reasoning: the highest propensity offer may be expensive or ineligible. The strategy filters ineligible offers, computes expected value as propensity times margin minus cost, and applies a contact-policy cap. The 0.6 offer wins on value, not propensity. The team monitors lift and override rate after deployment.
Traps
- Feature leakage: using post-interaction data that is unavailable at decision time.
- Training on all customers but scoring only eligible ones, causing skew.
- Treating adaptive model scores as calibrated probabilities without validation.
- Ignoring arbitration and contact policy, which can override model ranking.
- Deploying a model without a feedback loop and performance monitoring.