Goal: Use synthetic insurance data to simulate claims, discover risk profiles, and optimize pricing via unsupervised machine learning.
Tech Stack: Python, Pandas, Scikit-learn, Seaborn, PCA + KMeans
Key Outcome: 4 actionable customer clusters with pricing recommendations
License: MIT
This project demonstrates a full data science pipeline for insurance risk analytics:
- Raw synthetic customer data →
data/ - Monte Carlo claim simulation + clustering →
src/ - Visualizations + actionable insights →
results/
Results: 4 customer risk segments that reveal underpriced policies, high-risk losers, and profit drivers.
insurance-risk-segmentation/
│
├── data/ # Raw input data
│ ├── data_synthetic.csv
│ └── README.md # Dataset description
│
├── src/ # Analysis code
│ ├── analysis_pipeline.py
│ └── README.md # Step-by-step pipeline explanation
│
├── results/ # Outputs & insights
│ ├── cluster_summary.csv
│ ├── clustering_pca_2d.png
│ ├── frequency_histogram.png
│ ├── expected_loss_histogram.png
│ ├── severity_histogram.png
│ ├── freq_vs_sev_scatter.png
│ └── README.md # Full results & business actions
│
└── README.md # This file
| Cluster | Name | Risk | Avg. Premium | Expected Loss | Profit | Action |
|---|---|---|---|---|---|---|
| 0 | Gold Safe | Low | 3,223 € | 19.71 € | +3,203 € | Retain & upsell |
| 1 | High Frequency Loser | Very High | 3,539 € | 425 € | +3,114 € | Increase premium or exclude |
| 2 | Balanced Premium | Medium-High | 3,790 € | 147 € | +3,643 € | Maintain & reward |
| 3 | Claim Heavy, Low Premium | High | 1,939 € | 47.91 € | +1,892 € | Increase to 3.2k+ |
Critical Insight:
Cluster 3 has Claim History = 3.71 (near max) but pays only 1.9k → underpriced by ~60%.
| Plot | Insight |
|---|---|
| Clustering PCA 2D | Costumer clustering: 4 clear, compact groups → interpretable risk segments |
| Frequency distribution | Frequency distribution: Right-skewed most have 0–1 claims |
| Expected loss distribution | Loss distribution: Long tail - few high-cost customers dominate risk |
| Frequency vs Severity | Frequency vs Severity: Strong correlation; Risk Profile drives upper tail |
See all plots in results/
| Action | Estimated Benefit |
|---|---|
| Reprice Cluster 3 (+1.3k/customer) | +65 M€ (50k customers) |
| Exclude or penalize Cluster 1 | Reduce loss exposure |
| Upsell Cluster 0 & 2 | Increase retention & revenue |
# 1. Clone the repo git clone https://github.com/yourname/insurance-risk-segmentation.git cd insurance-risk-segmentation # 2. Install dependencies pip install pandas numpy matplotlib seaborn scikit-learn # 3. Run the analysis python src/analysis_pipeline.py
Outputs saved automatically to
results/
- Deterministic:
random_state=42in PCA, KMeans, and simulation - Pipeline-based: Consistent preprocessing
- Documented: Full READMEs in every folder
Developed by Salvatore Spagnuolo