Bank Telemarketing Propensity System

Telemarketing conversion is low and each call has real cost. The goal was to rank customers before outreach (without using post-call leakage signals like call duration) so teams can prioritize who to call first and measure lift in conversion and cost per acquisition.

• Defined the decision: pre-call ranking (what we know before calling) vs post-call outcomes (what happens during/after the call).
• Built leakage-free training data by excluding duration and other post-call signals; standardized preprocessing and evaluation.
• Trained and compared models (Logistic Regression, tree boosting, Neural Network) and validated with AUC + decile/capture metrics.
• Converted model scores into an operational call list: tiers/deciles + threshold guidance for different capacity levels.
• Added explainability (global + per-customer drivers) to support stakeholder trust and adoption.

• Enables a measurable pipeline from leads → ranked call list → tier-based outreach execution.
• Captures ~49% of subscribers in the top 10% of contacts and ~72% in the top 20% (capacity-friendly targeting).
• Estimated cost per acquisition improves ~€44 → ~€13 (~70%), with ~€165k savings per ~41k-contact campaign (planning estimate).
• Improves adoption readiness via explainability + documented rollout, monitoring, and retraining plan.

• Integration point: CRM exports a daily lead list → scoring pipeline → ranked call list returned to the calling team.
• Rollout: A/B test (ranked list vs baseline) with clear success metrics (conversion/cost-per-acquisition) and a fixed evaluation window.
• Monitoring: track score drift, conversion drift, tier performance, and recalibrate thresholds monthly/quarterly.
• Governance: leakage guardrails (no duration), data quality checks, and documented retraining cadence.

• Target leakage: call duration inflates performance—explicitly excluded to keep results realistic for pre-call decisioning.
• Class imbalance: success rate is low, so thresholds and capture metrics matter more than accuracy.
• Operational constraints: calling capacity changes—system supports tiered targeting strategies.