Bank Marketing Campaign Optimization 🏦

A complete machine learning pipeline for optimizing bank telemarketing campaigns using cost-sensitive learning, achieving 81.1% customer capture rate while minimizing campaign costs to $0.516 per contact.

---

Project Overview

This project implements a data mining solution to predict term deposit subscriptions from telemarketing campaigns. By applying cost-sensitive optimization techniques, we develop a model that balances recall (capturing potential customers) against precision (avoiding wasted calls).

Key Results

Metric	Value	Description
Recall	81.1%	Captures 81% of potential customers
Cost per Contact	$0.516	Optimized using custom cost matrix
ROC-AUC	0.804	Strong discrimination ability
Optimal Threshold	0.34	Cost-optimized decision boundary
Model	Logistic Regression	Winner after comparative analysis

Business Impact

For a 10,000-customer campaign:

• Expected acceptors: 1,130 (11.3% base rate)
• Our model captures: 917 customers (81.1%)
• Naive baseline captures: 376 customers (33.3%)
• Improvement: +541 customers (+144% lift)

Revenue Impact: At $100 profit per subscription = $54,100 additional revenue per 10k contacts

---

Features

• Complete ML Pipeline: Data loading → EDA → Preprocessing → Modeling → Optimization → Evaluation
• 21 Professional Visualizations: Class distributions, correlations, model performance, feature importance
• Cost-Sensitive Learning: Custom cost matrix reflecting business priorities (FN:FP = 13.3:1)
• Model Comparison: Decision Tree vs Logistic Regression with hyperparameter tuning
• Interpretability: Feature importance analysis and actionable business recommendations
• Reproducible: Automated pipeline script + Jupyter notebook with detailed explanations
• Academic Report: 5,500-word report following scientific standards

---

Project Structure

bankml/
├── README.md                           # This file
├── requirements.txt                    # Python dependencies
├── notebook.ipynb                      # Main Jupyter notebook (primary deliverable)
├── run_all.py                          # Automated pipeline script
├── report.md                           # Academic report (~5,500 words)
├── report.docx                         # Word version for Google Docs
│
├── data/
│   └── bank-marketing.csv              # UCI Bank Marketing Dataset (41K records)
│
├── assets/                             # 21 visualization images
│   ├── 01_class_distribution.png       # Target distribution
│   ├── 05_economic_indicators.png      # Economic context analysis
│   ├── 12_duration_leakage_analysis.png # Data leakage explanation
│   ├── 17_cost_threshold_optimization.png # Cost optimization
│   ├── 18_dt_feature_importance.png    # Decision Tree features
│   ├── 20_lr_coefficients.png          # Logistic Regression coefficients
│   └── 21_roc_comparison_final.png     # Final model comparison
│
├── models/                             # Trained models and results
│   ├── best_decision_tree.pkl
│   ├── best_logistic_regression.pkl
│   └── final_results.json
│
└── docs/                               # Documentation
    ├── EXECUTION_PLAN.md               # Detailed methodology
    └── SUBMISSION.md                   # Assignment submission guide

---

Quick Start

Prerequisites

Python 3.8+
pip

Installation

# Clone repository
git clone https://github.com/yourusername/bankml.git
cd bankml

# Install dependencies
pip install -r requirements.txt

Running the Project

Option 1: View Results (No Execution)

# Open Jupyter notebook to see complete analysis with outputs
jupyter notebook notebook.ipynb

# View academic report
cat report.md
# or open report.docx in Word/Google Docs

Option 2: Reproduce Full Pipeline

# Run complete pipeline (5-10 minutes)
python run_all.py

# This will:
# - Load and preprocess data
# - Generate 21 visualizations (saved to assets/)
# - Train and optimize models (Decision Tree + Logistic Regression)
# - Perform cost-sensitive threshold optimization
# - Save models to models/
# - Print performance metrics

Option 3: Interactive Exploration

# Launch Jupyter and run cells step-by-step
jupyter notebook notebook.ipynb

---

Methodology

1. Exploratory Data Analysis

• Dataset: 41,188 telemarketing contacts from Portuguese bank (2008-2013)
• Features: 20 input variables (demographics, campaign info, economic indicators)
• Target: Term deposit subscription (11.3% positive class)
• Key Finding: Duration variable shows data leakage → excluded from models

2. Data Preprocessing

• Missing Values: Mode/mean imputation
• Feature Engineering:
  • Created was_contacted_before (binary from pdays)
  • Log transformations: campaign_log, previous_log
  • Dropped duration (data leakage)
• Encoding: One-hot encoding for categorical variables
• Scaling: StandardScaler for numerical features
• Split: 75/25 stratified train-test split

3. Model Training & Optimization

Baseline Models

• Decision Tree (Entropy): 33.3% recall
• Logistic Regression (L2): 64.4% recall ✓ Better baseline

Hyperparameter Tuning (GridSearchCV)

• 5-fold Stratified Cross-Validation
• Decision Tree: max_depth, min_samples_leaf, ccp_alpha
• Logistic Regression: C, penalty, solver
• Result: DT recall improved to 62.1%

Cost-Sensitive Optimization

Cost Matrix:

False Positive (unnecessary call): +1.5
False Negative (missed customer): +20.0
True Positive (successful sale): -5.0
True Negative (correct avoid): 0.0

Threshold Optimization: Swept 0.01-0.99 to find minimum expected cost

4. Final Results

Model	Stage	Recall	Cost	ROC-AUC
Decision Tree	Baseline	33.3%	-	0.763
Decision Tree	Tuned	62.1%	-	0.801
Decision Tree	Optimized	69.4%	0.552	0.801
Logistic Regression	Baseline	64.4%	-	0.804
Logistic Regression	Tuned	64.4%	-	0.804
Logistic Regression	Optimized	81.1%	0.516	0.804

Winner: Logistic Regression with threshold=0.34

---

Key Insights

Top Predictive Features

Decision Tree Importance:

1. nr.employed (Employment level): 67.4%
2. cons.conf.idx (Consumer confidence): 13.0%
3. was_contacted_before: 5.4%

Logistic Regression Coefficients:

• Most Positive: month_mar (+1.07), cons.price.idx (+0.77)
• Most Negative: emp.var.rate (-1.69), month_may (-0.72)

Business Recommendations

1. Economic Timing Strategy

   • Launch campaigns during stable employment periods
   • Monitor macroeconomic indicators (employment, confidence)
   • Expected impact: 15-20% improvement

2. Warm Lead Prioritization

   • Previous contact increases acceptance 10x
   • Implement relationship nurturing programs

3. Seasonal Concentration

   • Focus on March (highest positive coefficient)
   • Reduce May activity (negative coefficient despite high volume)
   • Secondary peaks: September, October, December

4. Contact Method

   • Cellular > Telephone (telephone shows -0.64 coefficient)
   • Invest in cellular database maintenance

5. De-emphasize Demographics

   • Economic context matters more than age/job
   • Simplify targeting logic

---

Dataset

Source: UCI Machine Learning Repository - Bank Marketing Dataset

Citation:

Moro, S., Cortez, P., & Rita, P. (2014).
A Data-Driven Approach to Predict the Success of Bank Telemarketing.
Decision Support Systems, 62, 22-31.

Features:

• Demographics: age, job, marital, education
• Campaign: contact, month, day_of_week, campaign, pdays, previous, poutcome
• Economic: emp.var.rate, cons.price.idx, cons.conf.idx, euribor3m, nr.employed
• Target: y (term deposit subscription: yes/no)

---

🛠️ Technologies

• Python 3.8+
• Data Processing: pandas, numpy
• Visualization: matplotlib, seaborn
• Machine Learning: scikit-learn
  • DecisionTreeClassifier
  • LogisticRegression
  • GridSearchCV
  • StandardScaler
• Model Persistence: pickle
• Documentation: Jupyter, Markdown

---

Documentation

• notebook.ipynb - Complete analysis with code, outputs, and explanations
• report.md - Academic report (~5,500 words)
• EXECUTION_PLAN.md - Detailed methodology and experimental design
• SUBMISSION.md - Assignment submission guide

---

Visualizations

All 21 visualizations are available in the assets/ folder:

Exploratory Data Analysis (12 images):

• Class distribution, numerical distributions, correlations
• Age analysis, duration leakage demonstration
• Economic indicators, job/month seasonality
• Previous outcome impact, pdays distribution

Model Evaluation (9 images):

• Baseline confusion matrices and ROC curves
• Tuned model performance
• Cost-threshold optimization curves
• Feature importance (DT) and coefficients (LR)
• Final ROC comparison