Driver Behaviour Clustering for Insurance Risk

Overview

This project models how an auto-insurance company could move beyond one-size-fits-all pricing by grouping 10,000 policyholders into behaviourally coherent risk segments. It covers problem framing, feature analysis, clustering, validation, segment interpretation, and a production-ready scoring flow for new drivers.

System Design

The project is designed as a decision-support system, not just a clustering notebook. It translates model outputs into practical pricing actions, including loyalty discounts, risk-adjusted surcharges, and telematics programme recommendations.

Key Highlights

• Unsupervised driver segmentation
• K-Means selected after algorithm comparison
• PCA-based visual exploration
• Cluster validation using Silhouette, Davies-Bouldin, and Calinski-Harabasz metrics
• Driver persona cards
• New driver prediction tool
• Pricing recommendations by risk segment
• Responsible AI considerations for insurance pricing

Results

• Dataset size: 10,000 drivers
• Final algorithm: K-Means
• Optimal clusters: 2
• Silhouette Score: 0.3235
• Davies-Bouldin Index: 1.2377
• Calinski-Harabasz Index: 5706.15
• Identified Cautious Commuters and High-Exposure Drivers as distinct pricing-relevant segments