Bank Telemarketing Propensity System (ML Delivery & Decision Support)

Telemarketing outreach conversion is low and each call incurs real operational cost. The goal was to build a leakage-free propensity scoring model to prioritize high-potential customers before calling, optimizing agent capacity and minimizing campaign cost.

• Directed product strategy and full SDLC delivery using CRISP-DM framework from problem definition to executive readout.
• Built leakage-free feature sets by strictly excluding post-call duration signals; standardized preprocessing and validation.
• Trained and compared classification models (XGBoost, Logistic Regression, Neural Networks) evaluated on capture@decile and AUC.
• Translated model probabilities into operational decile tiers with clear calling-capacity thresholds.
• Built interactive Tableau dashboards translating model performance and profitability drivers for non-technical stakeholders.

• Captured ~49% of subscribers in the top 10% of contacts and ~72% in the top 20%, dramatically improving outbound efficiency.
• Delivered estimated ~$190K in campaign cost avoidance through strategic capacity reallocation.
• Enabled executive and operational adoption through visual explainability and documented rollout guidelines.

• Designed CRM lead integration pipeline: raw export → scoring model → ranked tiered call list.
• Rollout plan with A/B testing strategy against baseline outreach and evaluation windows.
• Monitoring framework tracking score drift, conversion drift, and threshold recalibration cadence.

• Target leakage: call duration artificially inflates model accuracy; explicitly excluded to ensure real-world pre-call validity.
• Class imbalance addressed through probability calibration and decile-based ranking rather than arbitrary cutoff thresholds.