đź“‹ Project Overview
In large-scale insurance environments, accurately forecasting the ultimate cost of a claim at the point of lodgement is a critical operational hurdle. This project focuses on the development of a high-performance predictive pipeline to estimate the Pure Premium (Expected Annual Liability) using a dataset of over 670,000 motor insurance policies.
By synthesising multi-dimensional risk features—including actuarial risk scores, geographic density, and vehicle specifications—the model identifies high-signal liability indicators early in the claim lifecycle. This solution serves as a technical proof-of-concept for evidence-based decision-making and proactive risk oversight within complex insurance portfolios.
đź”— View Live Interactive Dashboard
đź“Š About the Dataset: freMTPL2
The foundation of this model is the French Motor Third-Party Liability (TPL) Insurance Claims dataset. It provides a comprehensive view of risk characteristics observed primarily over a one-year period.
- Scope: Data collected for 677,991 motor third-party liability policies.
- Structure: The data is partitioned into two distinct tables:
freMTPL2freq: Captures policy-specific risk features (age, region, vehicle age, etc.) and the number of claims recorded.freMTPL2sev: Records the specific cost (amount) for each individual claim.
- Integration: Both tables are linked via a unique Policy ID (
IDpol) to create a unified view of frequency and severity.
đź’Ľ The Business Problem
Insurance organisations often face "claims leakage" and financial volatility due to reactive reserve setting. Traditional manual reviews are resource-intensive and may miss non-obvious correlations in high-volume data.
Strategic Objectives:
- Early Intervention: Automatically flags high-liability "outlier" claims at the point of entry for specialised management.
- Reserve Optimisation: Provides data-driven insights to set accurate financial reserves, ensuring organisational liquidity and regulatory compliance.
- Regulatory Adherence: Ensures data integrity and reporting standards align with governance frameworks.
🛠️ Technical Methodology
Addressing Zero-Inflation with Tweedie Regression
Insurance data is inherently "zero-inflated," meaning the vast majority of policies result in zero claims, while a small fraction result in highly skewed, positive costs.
To solve this, I implemented a Tweedie Regressor (p=1.5), which is a compound Poisson-Gamma distribution. This approach allows for the simultaneous modelling of claim frequency and severity in a single unified framework, providing significantly higher accuracy than standard linear models for insurance pricing.
The "Automated Data Supply Chain"
- End-to-End Pipeline: Developed a modular Scikit-learn pipeline to handle automated data acquisition, cleaning, and transformation.
- Feature Engineering: Implemented a
ColumnTransformerfor the technical verification of both numerical (Scaling) and categorical (One-Hot Encoding) risk features. - Exposure Weighting: Trained the model using
Exposureas a sample weight to ensure predictions are proportional to the policy duration.
đź“‚ Repository Structure
├── app.py # Streamlit Dashboard (Strategic Storytelling Interface)
├── tweedie_model.pkl # Production-ready trained pipeline (serialised)
├── requirements.txt # Dependency manifest for automated deployment
├── README.md # Project narrative and business impact
└── Notebooks/ # Annotated deep-dives into EDA and model validation
🚀 Business Impact & ROI
- Technical Proof of Concept: Demonstrates the ability to build "data apps" that allow stakeholders to explore real-time risk profiles independently.
- Enhanced Risk Oversight: Moves the analytical framework from retrospective reporting to proactive forecasting.
- Actionable Insights: Translates complex actuarial indicators (e.g., Bonus/Malus scores and Inhabitant Density) into clear, concise financial recommendations for senior stakeholders.