Bank Term Deposit Subscription Prediction: Portfolio Project

Portfolio Highlight: End-to-end ML pipeline for predicting term deposit subscriptions using advanced feature engineering and a tuned Artificial Neural Network (ANN).

This project demonstrates how to build a robust, interpretable, and business-relevant machine learning solution for bank marketing. The workflow—from data cleaning to ANN optimization—mirrors real-world data science best practices and is designed for portfolio presentation.

Table of Contents

• Project Overview
• Dataset Deep Dive
• Project Structure
• Quick Start
• Installation
• Usage
• Technical Deep Dive
  • Data Processing
  • Feature Engineering
  • Feature Selection
  • Model Development
• Results and Analysis
• Rationale for Model Selection
• Conclusion & Future Work
• Contributing
• License

Project Overview

Goal: Predict which bank customers will subscribe to a term deposit using a modern ML pipeline, with a focus on business value and model interpretability.

Key Steps:

• Data cleaning & feature engineering
• Exploratory data analysis (EDA)
• Feature selection
• ANN model development & tuning
• Evaluation & business recommendations

Data Overview & Preprocessing

Dataset: 41,188 samples, 21 features (demographics, financials, campaign data). Target: y (term deposit subscription).

Sample Data:

"age";"job";"marital";"education";"default";"housing";"loan";"contact";"month";"day_of_week";"duration";"campaign";"pdays";"previous";"poutcome";"emp.var.rate";"cons.price.idx";"cons.conf.idx";"euribor3m";"nr.employed";"y"
56;"housemaid";"married";"basic.4y";"no";"no";"no";"telephone";"may";"mon";261;1;999;0;"nonexistent";1.1;93.994;-36.4;4.857;5191;"no"
57;"services";"married";"high.school";"unknown";"no";"no";"telephone";"may";"mon";149;1;999;0;"nonexistent";1.1;93.994;-36.4;4.857;5191;"no"
... (see full dataset)

Cleaning Steps:

• Removed duplicates
• Replaced "unknown" in categorical columns with mode
• Encoded categorical features (one-hot, cyclic for months)

Project Structure

.
├── data/                 # Raw and processed datasets
├── notebooks/            # Jupyter notebooks for exploration, modeling, and evaluation
├── docs/                 # Project documentation, reports, and presentations
├── screenshots/          # Visualizations for portfolio and reporting
├── README.md             # Project overview and guide
└── requirements.txt      # Python dependencies

Quick Start

1. Clone the repository:

   git clone https://github.com/imaddde867/Bank-Term-Deposit-Prediction.git
   cd Bank-Term-Deposit-Prediction

2. Set up your environment:

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate
   pip install -r requirements.txt

3. Launch the Jupyter Notebook:

   jupyter notebook notebooks/ML-Final.ipynb

   Explore the notebook for the full workflow and code.

Usage

The main workflow is in notebooks/ML-Final.ipynb:

• Data loading, cleaning, and EDA
• Feature engineering and selection
• ANN model building, tuning, and evaluation

See docs/ for reports and rationale.

Technical Deep Dive

Data Processing & Feature Engineering

• Removed duplicates, handled 'unknown' values, encoded categoricals (one-hot, cyclic for months)
• Scaled numerical features (StandardScaler, MinMaxScaler)
• Ordinal encoding for education

Feature Selection

• Correlation analysis to drop redundant features
• Random Forest for feature importance (top 19 features retained)

Model Development: Artificial Neural Network (ANN)

Why ANN? Artificial Neural Networks (ANNs) are powerful for capturing complex, non-linear relationships in high-dimensional data. Here, an ANN outperformed tree-based models in AUC and generalization.

Architecture:

• Input: 19 features
• 3 hidden layers (128, 64, 32 neurons), each with ReLU, BatchNorm, Dropout
• Output: 1 neuron (sigmoid)

Training:

• Early stopping, learning rate scheduling
• 5-fold cross-validation for robust accuracy
• Hyperparameter tuning (batch size, learning rate, dropout, optimizer)

Results and Analysis

Performance Metrics:

Metric	Class 0 (No)	Class 1 (Yes)
Precision	0.90	0.76
Recall	0.99	0.16
F1-score	0.94	0.27
Support	3636	482

Overall:

• Accuracy: 0.8961
• ROC AUC Score: 0.7777
• 5-fold CV Accuracy: 0.8992 (±0.0022)

Classification Report:

              precision    recall  f1-score   support

           0       0.90      0.99      0.94      3636
           1       0.76      0.16      0.27       482

    accuracy                           0.90      4118
   macro avg       0.83      0.58      0.61      4118
weighted avg       0.88      0.90      0.87      4118

Confusion Matrix:

[[3611   25]
 [ 403   79]]

Rationale for Model Selection

Why ANN?

• Handles complex, non-linear relationships in mixed data
• Outperformed tree-based models in AUC and generalization
• Regularization (Dropout, BatchNorm) and learning rate scheduling ensured stability
• Despite class imbalance, provided actionable leads for marketing
• Probability outputs allow for business-driven threshold tuning

Conclusion & Future Work

This project showcases a full ML pipeline for a real-world business problem, with a focus on ANN modeling and interpretability. The approach is generalizable to other imbalanced, high-dimensional business tasks.

Next Steps:

• Explore advanced imbalance techniques (SMOTE, class weights)
• Try ensemble models
• Further feature engineering
• Adjust classification threshold for business needs

License

MIT License. See LICENSE.