Project 1 | Predictive Engagement Modeling for Retail Banking

Data science is about transforming raw data into actionable insights that inform better decisions. In this project, I applied my skills in data cleaning, predictive modeling, and data visualization to analyze customer engagement patterns in a banking context. This project demonstrates my ability to preprocess large datasets, construct and validate machine learning pipelines, and create clear, interpretable dashboards. In real-world applications, this work helps businesses identify at-risk customers, optimize resource allocation, and reduce churn—leading to improved efficiency, customer retention, and strategic decision-making. 💡📊🤖🏦


Links 🔗

Presentation 🎤 My GitHub 💻 Team Repo 👥

Project Overview 📄

This analysis transforms the UCI Bank Marketing dataset (41,188 records from Portuguese bank telemarketing campaigns, 2008-2010) from a term deposit subscription prediction task into a customer churn prevention framework. By reframing "no" responses as disengagement signals, we identify at-risk customers and provide actionable strategies to shift from inefficient mass calling to targeted retention efforts.


Project Workflow 🪜

Visuals 📷


Dashboard: Soft churn visualization

Dashboard: Soft churn visualization

Figure 1: This dashboard visualizes the baseline 88/12 class imbalance and identifies “soft churn” customers through duration analysis. It allows management to see that disengagement is often visible early in the call cycle.

KDE density plot

KDE density plot

Figure 2: This KDE density plot highlights the concentration of likely disengagement. Business Action: The dark “Red Zone” identifies customers aged 30-50 with calls under 200 seconds. Representatives should be trained to conclude these low-propensity calls quickly to save resources.


Results 🟰

Business Impact 👥: Model achieves 89.1% accuracy and 49.5% precision when targeting top 10% highest-probability customers (vs. 11.3% baseline conversion).

Critical Risk Zone 🛑: Middle-aged customers (30-50) with calls <200 seconds during unfavorable economic conditions.

Key Findings 🔑: Top 10 Churn Drivers include duration (call length), euribor3m, age, nr.employed, campaign, pdays_never_contacted, emp.var.rate, cons.price.idx, cons.conf.idx, pdays.


Technical Stack 🔨

Core: pandas, scikit-learn, matplotlib, seaborn

Imbalance Handling: imblearn (SMOTE)

Preprocessing: ColumnTransformer, StandardScaler, OneHotEncoder

Models: LogisticRegression, RandomForestClassifier

Evaluation: ROC-AUC, Precision-Recall curves, confusion matrices


Ethical Considerations ⚖️