Insurance Risk Analytics & Predictive Modeling for ACIS

This project explores and models risk and profitability in insurance data using statistical reasoning, EDA, version control with DVC, and machine learning techniques. It forms the foundation for a data-driven, risk-based premium pricing system.

---

Project Objective

To analyze insurance policy and claims data to:

• Understand customer risk and claim patterns (EDA)
• Quantify risk using loss ratio and other metrics
• Statistically validate segmentation hypotheses
• Predict claim severity, claim occurrence, and optimize premiums
• Enable transparency using explainable ML (SHAP)

---

Project Structure

Insurance-Risk-Analytics---Predictive-Modeling-for-ACIS/
├── data/                         # Raw & processed datasets (DVC-managed)
│   ├── machineLearningRating_v3.txt
│   ├── cleaned_machineLearningRating.csv
│   ├── processed_for_modeling.csv
│   └── cleaned_machineLearningRating.csv.dvc
├── src/                          # Modularized Python code
│   ├── data_load.py              # Load & clean dataset
│   ├── eda.py                    # EDA functions & plots
│   └── __init__.py                 # Reusable helpers
├── notebooks/                    # Jupyter notebooks
│   ├── import_clean.ipynb        # Cleaning stage
│   ├── eda.ipynb                 # EDA & visual insights
│   ├── task3\_hypothesis.ipynb    # A/B Hypothesis Testing
│   └── task4\_modeling.ipynb      # Predictive Modeling
|   ├── plots/                        # Saved plots (grouped by type)
|    │   ├── univariate/
|    │   ├── bivariate/
|    │   ├── multivariate/
|    │   ├── time/
|    │   ├── outliers/
|    │   ├── geography/
|    │   └── vehicle_risks/
|    ├── saved_models/                      
|    │   ├── randomforest_best_model.pkl
|    │   ├── scaler.pkl
|    │   ├── sgb_claim_occured_model.pkl
|    │   ├── xgboost_severity_model.pkl
|
├── APP.PY
├── .gitignore
├── .dvcignore
├── README.md
├── LICENSE
└── requirements.txt

---

Task 1: Exploratory Data Analysis (EDA)

Objectives

• Understand insurance risk patterns
• Assess data quality and completeness
• Explore geographic and demographic patterns

Key Analyses Performed

• Data Summarization: Descriptive statistics, data types, skewness
• Missing Value Assessment: Titles used to infer missing Gender and MaritalStatus
• Feature Engineering: Created LossRatio = TotalClaims / TotalPremium
• Univariate & Bivariate Analysis: Histograms, bar charts, segmented loss ratios
• Temporal Trends: 18-month premium/claim patterns
• Outlier Detection: Boxplots of financial metrics
• Top Risky Vehicles: Highest claim make and model
• Loss Ratio by Zip Code: Geographic risk mapping

All visualizations are saved under plots/.

---

Task 2: Data Versioning with DVC

Purpose

Enable reproducible data pipelines with version-controlled datasets.

DVC Workflow Completed:

• dvc init initialized in repo
• Created a local remote: ~/dvc-storage
• Tracked cleaned dataset:
  dvc add data/cleaned_machineLearningRating.csv
• Committed .dvc files and .gitignore updates
• dvc push successfully pushed to local remote

This guarantees traceability, reproducibility, and collaboration readiness.

---

Task 3: A/B Hypothesis Testing

Goal: Statistically validate risk & margin segmentation hypotheses.

Hypothesis #	Null Hypothesis	Result
H₀₁	No risk difference across Provinces	Rejected
H₀₂	No risk difference across Zip Codes	Rejected
H₀₃	No margin difference across Zip Codes	Rejected
H₀₄	No risk difference between Genders	Failed to Reject
Tracking Device	No risk difference due to Tracking Device	Failed to Reject
NewVehicle	No risk difference between New vs Old vehicles	Rejected

• Tests Used: Chi-squared, ANOVA, Z-test (for proportions)
• KPIs: Claim Frequency, Claim Severity, Profit Margin

---

Task 4: Predictive Modeling

1. Claim Severity Model (Regression)

• Target: TotalClaims for rows where TotalClaims > 0
• Models: Linear Regression, Decision Tree, Random Forest, XGBoost
• Best Performance:
  • XGBoost: RMSE = ~14,257, R² = ~0.87

2. Claim Probability Classifier (Binary Classification)

• Target: Whether a claim occurred
• Metrics: Accuracy, Precision, Recall, F1-score, ROC AUC
• Best Model:
  • XGBoost: ROC AUC = 0.901

SHAP (Explainability)

• SHAP used to interpret XGBoost outputs for:
  • Claim Severity Regression
  • Claim Classification (Probability)
• Top Drivers Identified: make, VehicleType, Province, SumInsured, etc.

---

Key Learnings & Business Value

• Provinces & Zip Codes show significant risk variation → segment pricing by geography
• Some vehicle types/models are disproportionately risky
• XGBoost consistently performs best, but business interpretability must be balanced
• SHAP explanations provide model transparency → crucial for insurance

---

Next Steps

• Tune models with cross-validation & hyperparameter optimization
• Integrate into Streamlit dashboard (Task 5 - Optional Bonus)
• Deploy risk-based pricing system to provide fair premiums

Sample Outputs

Dashboard Dashboard Dashboard

---

Dependencies

Install required packages:

pip install -r requirements.txt

---

Author

• Nurye Nigus
• Project: Insurance Risk Analytics & Modeling for ACIS

---