Health Insurance Prediction – Machine Learning Project

Overview

This project focuses on predicting whether a customer has health insurance using demographic, financial, and lifestyle data.

The goal is to support data-driven marketing strategies, allowing companies to target high-potential customers, reduce costs, and improve customer acquisition.

---

Business Problem

Marketing campaigns for health insurance are often broad and inefficient.

This project answers the key question:

How can we predict which customers are most likely to subscribe to health insurance in order to optimize marketing efforts and improve inclusion?

---

Dataset

• Observations: 72,458 customers
• Features: 15 variables
• Target: health_ins (1 = Has insurance, 0 = No insurance)

Feature Types:

• Demographic → age, sex, marital status
• Financial → income
• Lifestyle → housing, vehicles, gas usage
• Geographic → state of residence

---

Key Insights from Analysis

• Income is the strongest predictor of insurance ownership
• Older individuals are more likely to have insurance
• More vehicles and better housing → higher probability of insurance
• Gas usage showed low predictive power
• Strong class imbalance (majority already insured)

---

Data Preparation

• Missing values handled (median / mode imputation)

• Outliers treated (capping + log transformation)

• Feature scaling (Standardization)

• Categorical encoding (One-Hot Encoding)

• Feature engineering:

  • income_per_vehicle
  • income_age_interaction
  • financial_status
  • mobility_indicator

---

Models Implemented

• Logistic Regression
• Random Forest
• XGBoost
• Gradient Boosting
• Neural Network (MLP)

Model Performance

Model	Accuracy	F1-Score	AUC
Logistic Regression	0.68	0.79	0.76
Random Forest	0.84	0.91	0.73
XGBoost	0.69	0.80	0.79
Gradient Boosting	0.90	0.95	0.80
Neural Network	0.23	0.27	0.47

---

Best Model

Gradient Boosting achieved the best overall performance:

• High accuracy (0.90)
• Strong balance between precision and recall
• Robust handling of class imbalance

---

Business Impact

This model enables:

• Targeted marketing campaigns
• Reduction in marketing costs
• Higher conversion rates
• Better understanding of customer behavior
• Improved inclusion of underserved groups

---

Model Deployment

• Final model saved using joblib
• Predictions generated on unseen test data
• Output file: final_predictions.csv

---

Key Learnings

• Feature engineering has a major impact on performance
• Handling class imbalance is critical
• Simpler models can be strong baselines
• Ensemble methods (Boosting) perform best on structured data

---

Tech Stack

• Python
• Pandas, NumPy
• Scikit-learn
• XGBoost
• Matplotlib, Seaborn

---

Conclusion

This project demonstrates how machine learning can be applied to solve a real-world business problem by combining data analysis, feature engineering, and predictive modeling.

The final model provides both accurate predictions and actionable insights, making it valuable for strategic decision-making.