🌐 English · Español

🏦 Bank Marketing Campaign Analysis

→ Try the interactive demo — score any client profile, explore segments, and review model performance.

Analyzing direct marketing campaigns of a Portuguese banking institution to predict whether a client will subscribe to a term deposit. This project combines exploratory data analysis, statistical testing, customer segmentation, and predictive modeling to derive actionable business insights.

---

🎯 Objective

Business Question: Which client profiles are most likely to subscribe to a term deposit, and how can the bank optimize its campaign strategy to increase conversion rates?

This project addresses the question from multiple angles:

• Descriptive: Who are the clients that subscribe? What campaign patterns drive conversion?
• Inferential: Are conversion rate differences across segments statistically significant?
• Predictive: Can we build a model to identify high-probability subscribers before calling them?
• Prescriptive: What actionable recommendations can improve campaign ROI?

---

📊 Dataset

Feature	Detail
Source	UCI Machine Learning Repository
Reference	Moro et al. (2014). A Data-Driven Approach to Predict the Success of Bank Telemarketing
Records	45,211
Features	16 input + 1 target
Target	y — Has the client subscribed a term deposit? (yes/no)
Class Balance	88.7% No / 11.3% Yes (imbalanced)

Key variables: age, job, marital status, education, balance, housing loan, contact type, campaign duration, number of contacts, days since previous contact, previous campaign outcome.

⚠️ Note on duration: Per the UCI documentation and Moro et al. (2014), the duration field is only known after a call ends and is therefore excluded from all predictive models to avoid data leakage.

---

🔬 Methodology

The analysis is structured as a Databricks pipeline with three notebooks:

notebook_1_EDA.ipynb              notebook_2_Limpieza_Preprocesamiento.ipynb
  ├── Schema & quality checks       ├── Encoding (Label / OHE)
  ├── Univariate analysis           ├── Feature engineering
  ├── Bivariate analysis            ├── Stratified train/test split
  ├── Conversion-rate segments      └── Persisted as parquet
  └── Statistical testing

notebook_3_MachineLearning.ipynb
  ├── Decision Tree (baseline)
  ├── Random Forest + GridSearch
  ├── XGBoost + GridSearch
  ├── Stratified K-Fold CV
  └── ROC-AUC evaluation on hold-out test set

Stack: PySpark for data processing, scikit-learn + XGBoost for modeling, matplotlib/seaborn for visualization.

---

📈 Key Results

Model Performance (Hold-out Test Set, ROC-AUC)

Model	CV AUC	Test AUC	Notes
Decision Tree (max_depth=5)	—	0.7745	Interpretable baseline
XGBoost (default)	—	0.7819	Strong out-of-the-box
XGBoost (GridSearch)	0.9611	0.7647	⚠️ CV-test gap reveals leakage
Random Forest (best)	—	0.7959 ★	Best generalization

Critical finding: The XGBoost grid-search CV reported AUC ≈ 0.96 but only achieved 0.76 on the held-out test set. Root cause: SMOTE was applied to the full training set before CV, leaking synthetic samples across folds. The fix — applying SMOTE inside an imblearn.Pipeline so each fold resamples independently — is implemented in the v2 iteration. This is exactly the kind of methodology bug a real production pipeline must catch.

Main Findings

1. Severe class imbalance (88.7% / 11.3%) — accuracy is misleading; ROC-AUC and PR-AUC are the relevant metrics, and resampling must be CV-aware.
2. Previous contact is the strongest signal — clients previously contacted convert at 63.8% vs 9.3% for new clients (≈ 7× lift). Re-engagement campaigns should be prioritized.
3. High-converting segments: students (31%) and retirees (25%) convert at roughly 3× the base rate (11.3%). Both segments are systematically under-targeted in the current campaign mix.
4. Channel inefficiency in May: 13,769 calls placed but only 6.4% converted — the worst ROI of the year. Volume-based scheduling outperformed by quality-based targeting.
5. Top predictive features (Random Forest importance): poutcome_success, pdays, previous, month, contact_type, balance.

---

🗂️ Project Structure

bank-marketing-analysis/
├── README.md
├── requirements.txt
├── .gitignore
├── notebook_1_EDA.ipynb
├── notebook_2_Limpieza_Preprocesamiento.ipynb
└── notebook_3_MachineLearning.ipynb

The databricks-version branch contains the original Databricks .py source format for direct import into a workspace.

---

🔁 v2 Iteration

A second version of the modeling pipeline lives on the v2-imblearn-pipeline branch. It addresses the SMOTE-in-CV leakage diagnosed above by wrapping resampling and the classifier in an imblearn.pipeline.Pipeline, plus adds:

• Threshold tuning optimized for F1 / business cost
• PR-AUC alongside ROC-AUC (more honest under heavy imbalance)
• joblib model + encoder persistence
• Pinned requirements.txt for reproducibility

---

🛠️ Tech Stack

Python PySpark Pandas NumPy Scikit-learn XGBoost imbalanced-learn Matplotlib Seaborn

---

🚀 How to Reproduce

# 1. Clone the repository
git clone https://github.com/jsanchez-ds/bank-marketing-analysis.git
cd bank-marketing-analysis

# 2. Install dependencies
pip install -r requirements.txt

# 3. Download the dataset from UCI
#    https://archive.ics.uci.edu/dataset/222/bank+marketing

# 4. Run the notebooks in order
jupyter notebook notebook_1_EDA.ipynb

To run on Databricks instead, import the .py files from the databricks-version branch.

---

---

Live Demo (Streamlit)

→ Try it live

An interactive Streamlit app wraps the Random Forest classifier and exposes it as a scoring tool with three tabs:

• Predict — score a hypothetical client with a tunable decision threshold and CALL/SKIP recommendation.
• Insights — headline conversion-rate findings (warm leads ~7x lift, students ~3x base rate, May campaign ROI).
• Model Card — held-out ROC-AUC / PR-AUC, ROC curve, confusion matrix, top 15 features, and the duration leakage caveat.

Run locally:

pip install -r app/requirements.txt
python app/train_model.py            # one-time, ~30s
streamlit run app/streamlit_app.py

See app/README.md for the full guide and Streamlit Community Cloud deployment steps.

---

👤 Author

Jonathan Sánchez

• GitHub: @jsanchez-ds
• Universidad de Chile — Industrial Engineering