Bank Marketing Machine Learning Project: Data Science Practice from Loss to Profit

The BYU-Idaho CSE450 course team project addresses the class imbalance problem in bank telemarketing (only 11.4% positive cases) by using SMOTE oversampling, class weights, probability threshold tuning, and multi-model comparison (random forest, ensemble stacking, etc.), turning marketing campaigns from loss to profit while generating interpretable customer insights.

Bank telemarketing faces problems of low efficiency and high cost, especially class imbalance (10-15% positive cases) which makes traditional models have no business value. The project uses the UCI Bank Marketing Dataset (about 37,000 records, 11.4% positive cases), with features covering customer basic information, financial status, contact history, and macroeconomic indicators.

The team built three models: 1. Random Forest + SMOTE oversampling + manual class weights; 2. Balanced Random Forest (automatic class weights); 3. Ensemble Stacking (RF + KNN base learners + Logistic Regression meta-learner, threshold tuned to 0.61). Class imbalance handling techniques include SMOTE interpolation to generate synthetic samples, class weight adjustment, and probability threshold optimization.

Without ML screening, the test set lost $157; after the best model, it made a profit of $824, and scaling to 4119 people is expected to yield a profit of $7775. Customer insights: High conversion groups (historical converted customers, students, retirees), low conversion groups (fixed-line customers, blue-collar workers), precision rate increased from 11.5% to 47.2%.

Key success factors: Business problem-driven (focus on profit goals), choose business-relevant evaluation metrics (not accuracy), emphasize interpretability (generate customer insights), iterative optimization (from simple to complex models).

Model aspect: Deep learning, time series, reinforcement learning; Business aspect: Personalized recommendation, contact timing optimization, customer lifetime value prediction; Technology aspect: Real-time inference, A/B testing, model monitoring.