Bank Marketing — Random Forest (Calibrated)

Predicting Term-Deposit Subscription from the UCI “Bank Marketing” dataset

---

📌 Executive Summary

This project develops a calibrated Random Forest classifier to predict whether a client will subscribe to a term deposit following a direct marketing campaign. The workflow emphasizes:

• Feature engineering for marketing history & socio-economic signals
• Stratified cross-validation with random/grid search tuning
• Probability calibration for trustworthy decision thresholds

Key Outcomes (Test Split 25%):

• Accuracy: 0.851
• ROC-AUC: 0.817
• PR-AUC (Avg Precision): 0.496
• Confusion Matrix: [[8014, 1123], [413, 747]]
• Positive class (subscribe): Precision 0.40 | Recall 0.64 | F1 0.49

⚡ In a highly imbalanced setting (~11.3% positives), recall prioritization ensures more true subscribers are identified, which is crucial for campaign ROI.

---

📊 Dataset Overview

• Source: UCI Bank Marketing Dataset
• Shape: 41,188 rows × 21 columns
• Target: y (converted into binary target)
• Class Balance: 11.3% yes, 88.7% no

Target Distribution:

Job vs Target:

Campaign count vs Target:

Euribor 3-month rate vs Target:

🔎 The dataset contains “unknown” values in default, education, housing, loan, etc. The duration variable was removed to avoid data leakage.

---

🛠️ Approach & Pipeline

🔹 Feature Engineering

• Dropped duration (leakage)

• Created marketing history indicators:

  • was_prev_contacted (binary)
  • pdays_when_contacted (numeric, NaN for never contacted)

• Engineered binary target variable

🔹 Preprocessing

• Categorical: OneHotEncoder
• Numerical: Pass-through

🔹 Model

• RandomForestClassifier(class_weight='balanced', random_state=42, n_jobs=-1)
• Integrated in Pipeline with preprocessing

🔹 Train/Test Split

• Train: 30,891 rows
• Test: 10,297 rows

---

⚙️ Hyperparameter Tuning & Calibration

• RandomizedSearchCV (12k subsample) + GridSearchCV
• Optimized for ROC-AUC
• Final model wrapped in CalibratedClassifierCV (isotonic, cv=5)

Performance after Calibration:

• ROC-AUC: 0.817
• PR-AUC: 0.496
• Brier Score: 0.074

ROC Curve:

---

✅ Results (Test Set)

Classification Report:

Class 0: precision 0.95 | recall 0.88 | f1 0.91 | support 9137
Class 1: precision 0.40 | recall 0.64 | f1 0.49 | support 1160
Overall accuracy: 0.851 | ROC-AUC: 0.817 | PR-AUC: 0.496

Confusion Matrix:

---

🔑 Feature Importance & Explainability

Top Global Feature Importances:

1. nr.employed
2. euribor3m
3. emp.var.rate
4. cons.conf.idx
5. age
6. pdays_when_contacted
7. cons.price.idx
8. campaign
9. month_may, month_oct
10. was_prev_contacted

SHAP Analysis:

• Global importance:
• Impact distribution:
• Individual predictions:

---

📈 Business Framing & Impact

• Lead scoring: prioritize clients most likely to subscribe
• Campaign efficiency: maximize ROI by targeting likely converters
• Imbalance-aware: recall-oriented thresholding captures more subscribers
• Explainability: SHAP-based insights enhance trust and compliance

---

⚡ How to Reproduce

Setup

python -m venv .venv && source .venv/bin/activate
pip install -U scikit-learn==1.5.1 pandas numpy matplotlib seaborn joblib

Run

1. Place bank-additional-full.csv in ./data/
2. Run notebook → trains, tunes, and calibrates the model
3. Artifacts are saved in ./models and ./outputs

---

📂 Project Structure

.
├── notebooks/
│   └── Random forest tree banking.ipynb
├── models/
│   ├── rf_bank_marketing_calibrated.joblib
│   └── encoded_feature_names.csv
├── outputs/
│   └── feature_importances_top25.csv
├── images/
│   ├── Target count.png
│   ├── job vs target.png
│   ├── campaing count by target.png
│   ├── euribor 3 month rate density vs target.png
│   ├── R0o curve.png
│   ├── confusin metriucs.png
│   ├── Feature inmportances.png
│   ├── shap va;ues importances.png
│   ├── shap va;ue prediction imoact.png
│   ├── shap values impact on output.png
├── data/
│   └── bank-additional-full.csv  (not tracked)
├── README.md
└── requirements.txt

---

⚖️ Governance, Risks & Next Steps

• Fairness: Age & employment-related features → monitor bias
• Concept drift: Macro variables → retrain regularly
• Data quality: Handle “unknown” values in defaults/education
• Future Work: Gradient boosting, cost-sensitive learning, real-time REST API

---

📌 Attribution

• Dataset: UCI Bank Marketing
• Author: Steffin Thomas
• Environment: Python 3.12.7, scikit-learn 1.5.1
• Notebook Run: Aug 2025

---

🚀 Highlights

• Clean & reproducible end-to-end ML pipeline
• Balanced handling of class imbalance & calibration
• Strong business framing + explainability for recruiter review
• Deployment-ready artifacts

---