🚗 Car Insurance Claim Prediction

Python scikit-learn XGBoost License

A comprehensive binary classification analysis predicting insurance claims across two distinct datasets: vehicle specifications (imbalanced) and driver profiles (balanced)

View Vehicle Analysis View Driver Analysis Dataset Source1 Dataset Source2

📋 Table of Contents

🎯 Project Overview

This project tackles the car insurance claim prediction problem from two complementary angles: vehicle/equipment characteristics and driver behavioral profiles. By analyzing both datasets, we provide insurers with actionable insights into risk factors that span both the insured asset and the policyholder’s driving history.

Business Problem

Insurance companies need accurate claim prediction models to:

However, claim datasets often suffer from severe class imbalance (only 6-7% of policies result in claims), making traditional accuracy metrics misleading and requiring specialized modeling techniques.

Objectives

✨ Key Features

Dual Dataset Analysis

Advanced Preprocessing Pipeline

Comprehensive EDA & Feature Engineering

Model Development & Tuning

📊 Datasets

Dataset 1: Vehicle/Equipment Specifications (Imbalanced)

Source: Kaggle Insurance Claims Dataset
Size: 58,592 records → 58,323 after outlier filtering
Target: claim_status (≈6.4% positive class)
Features: 41 original columns including:

Feature Category Examples
Policy Info subscription_length, region_code, region_density
Vehicle Specs vehicle_age, segment, fuel_type, displacement, cylinder
Safety Features airbags, ncap_rating, is_esc, is_tpms, is_brake_assist
Equipment Flags is_parking_camera, is_central_locking, is_power_steering
Engine Details max_torque, max_power, engine_type, transmission_type

Note: Many equipment flags showed near-perfect correlation (Cramér’s V > 0.95) and were pruned

Dataset 2: Driver Profile & Behavior (Balanced)

Source: Kaggle Car Insurance Data
Size: 10,000 records
Target: OUTCOME (31.3% positive class)
Features: 19 original columns including:

Feature Category Examples
Demographics AGE, GENDER, RACE, EDUCATION, INCOME, MARRIED, CHILDREN
Financial CREDIT_SCORE (9.8% missing)
Driving Profile DRIVING_EXPERIENCE, ANNUAL_MILEAGE (9.6% missing)
Vehicle Info VEHICLE_OWNERSHIP, VEHICLE_YEAR, VEHICLE_TYPE
Risk Indicators SPEEDING_VIOLATIONS, DUIS, PAST_ACCIDENTS
Geographic POSTAL_CODE (high cardinality → target encoded)
Purchase BOOKING_SOURCE

Note: GENDER and RACE excluded from training to avoid discriminatory modeling

🔬 Technical Approach

1. Data Preprocessing Strategy

Common Pipeline (Both Datasets)

# Preprocessing order (critical for avoiding leakage):
1. Drop identifiers and redundant columns
2. Handle outliers (Tukey method for vehicle_age)
3. Train-test split (stratified by target)
4. Target encoding for high-cardinality categoricals (cross-fitted)
5. One-Hot Encoding for remaining categoricals
6. Imputation and scaling in sklearn Pipeline

Dataset-Specific Considerations

Vehicle Dataset:

Driver Dataset:

2. Feature Engineering Highlights

Target Encoding (Cross-Fitted)

def target_encode_crossfit(X, y, col, n_splits=5, smoothing=50):
    # Stratified K-Fold to prevent leakage
    # Smoothed mean: (count × mean + smoothing × global) / (count + smoothing)
    # Produces numeric risk score per category

Correlation-Driven Pruning

3. Modeling Strategy

We evaluated multiple model architectures with class-imbalance handling across both datasets:

Baseline: Logistic Regression

Ensemble: Random Forest

Gradient Boosting: XGBoost (Vehicle Dataset Only)

4. Threshold Selection via Precision-Recall Curve

For imbalanced problems, the default 0.5 threshold is often suboptimal. We analyzed:

  1. Default (0.5): Baseline confusion matrix
  2. Best F1: Threshold maximizing F1-score on PR curve
  3. Precision-constrained: Threshold achieving minimum precision (e.g., 0.25) with maximum recall 
    prec, rec, thr = precision_recall_curve(y_test, probs)
f1_vals = 2 * prec[:-1] * rec[:-1] / (prec[:-1] + rec[:-1] + 1e-12)
best_threshold = thr[f1_vals.argmax()]

🏆 Results

Dataset 1: Vehicle/Equipment Specifications (Imbalanced)

Model Performance Comparison

Model ROC-AUC PR-AUC Accuracy @ 0.5 Log Loss Notes
Logistic Regression 0.617 ~0.090 0.559 0.675 Baseline linear model
Random Forest 0.660 0.109 0.732 0.517 Best tuned ensemble ✅
XGBoost 0.669 0.115 0.582 0.632 Highest ROC-AUC but low precision

Best Model: Random Forest (Final)

Configuration:
- n_estimators: 300
- max_depth: 12
- min_samples_leaf: 10
- max_features: 'sqrt'
- class_weight: 'balanced'

Performance @ threshold = 0.5:
✅ ROC-AUC: 0.660
✅ PR-AUC: 0.109
✅ Accuracy: 0.732
✅ Precision: 0.112 (112 claims per 1000 flagged)
✅ Recall: 0.461 (catches 46% of true claims)
✅ F1-Score: 0.181

Confusion Matrix (Test Set: 11,665 samples):
              Predicted
              No    Yes
Actual No   8194   2723
       Yes   403    345

Key Findings (Vehicle Dataset)

  1. Severe imbalance limits precision
    • Only 6.4% positive class makes high precision difficult
    • Trade-off: Catching 46% of claims requires flagging ~27% of policies
  2. Tree ensembles outperform linear models
    • RF achieved 7% higher ROC-AUC vs. Logistic Regression
    • Non-linear interactions (e.g., region × vehicle_age × segment) matter
  3. Equipment redundancy was massive
    • 20+ features dropped due to Cramér’s V > 0.90
    • Example: is_power_steering perfectly predicted by segment
  4. XGBoost showed highest discrimination but lower precision
    • Best ROC-AUC (0.669) but struggled with precision at useful recall
    • RF better balanced precision-recall trade-off

Dataset 2: Driver Profile & Behavior (Balanced)

Model Performance Comparison

Model ROC-AUC PR-AUC Accuracy @ 0.5 Precision @ 0.5 Recall @ 0.5 F1-Score
Logistic Regression 0.906 0.812 0.842 0.755 0.732 0.743 ✅
Random Forest 0.897 0.787 0.829 0.721 0.742 0.731

Best Model: Logistic Regression

Configuration:
- class_weight: 'balanced'
- max_iter: 2000
- solver: 'lbfgs'

Performance @ threshold = 0.5:
✅ ROC-AUC: 0.906 (excellent discrimination)
✅ PR-AUC: 0.812 (strong ranking)
✅ Accuracy: 0.842
✅ Precision: 0.755 (75.5% of flagged cases are true claims)
✅ Recall: 0.732 (catches 73% of claims)
✅ F1-Score: 0.743

Confusion Matrix (Test Set: 2,000 samples):
              Predicted
              No    Yes
Actual No   1224    149
       Yes   168    459

Key Findings (Driver Dataset)

  1. Balanced data enables strong performance
    • 31% positive class allows much better precision (75.5% vs. 11.2%)
    • Both models achieve >82% accuracy
  2. Linear model wins on balanced data
    • Logistic Regression outperforms RF across all metrics
    • Driver risk appears approximately linear in feature space
    • Simpler model → better interpretability for stakeholders
  3. Precision-recall trade-off is favorable
    • LR: 149 false positives, 168 false negatives
    • RF: 180 false positives, 162 false negatives
    • LR finds 6 fewer true claims but creates 31 fewer false alarms
  4. Target-encoded POSTAL_CODE was crucial
    • Geographic risk patterns captured without sparse OHE
    • Cross-fitting prevented overfitting

Cross-Dataset Insights

Aspect Vehicle Dataset (Imbalanced) Driver Dataset (Balanced)
Best Model Random Forest Logistic Regression
Class Imbalance Impact Severe (6.4% positive) → low precision unavoidable Moderate (31% positive) → high precision achievable
Feature Space Non-linear (equipment × region interactions) Approximately linear (additive risk factors)
ROC-AUC 0.660 (moderate discrimination) 0.906 (excellent discrimination)
Precision @ 0.5 0.112 (poor) 0.755 (strong)
Business Action Need higher threshold or cost-benefit analysis Can flag high-risk drivers confidently

Model Comparison Visualization

ROC-AUC Performance:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Vehicle Dataset (Imbalanced):
  Logistic Reg   ████████████        0.617
  Random Forest  █████████████       0.660 ⭐
  XGBoost        █████████████       0.669

Driver Dataset (Balanced):
  Logistic Reg   ██████████████████  0.906 ⭐
  Random Forest  █████████████████   0.897

Precision @ 0.5:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Vehicle Dataset:
  Best (RF)      ██                  0.112

Driver Dataset:
  Best (LR)      ███████████████     0.755

🚀 Installation & Usage

Prerequisites

Python 3.8+
pip install -r requirements.txt

Dependencies

pandas>=1.3.0
numpy>=1.21.0
scikit-learn>=1.0.0
xgboost>=1.5.0
matplotlib>=3.4.0
seaborn>=0.11.0
scipy>=1.7.0
joblib>=1.0.0

Quick Start

# Clone the repository
git clone git@github.com:ROYBRUNO81/insurance.github.io.git
cd car-insurance-claim-prediction

# Install dependencies
pip install -r requirements.txt

# Download datasets from Kaggle (see Dataset section)
# Place in project root:
#   - insurance_claims.csv
#   - Car_Insurance_Claim.csv

# Launch Jupyter Notebook
jupyter notebook

Running the Analyses

Vehicle Dataset Analysis

# Open Portfolio_Project.ipynb
1. Load data (insurance_claims.csv)
2. Run EDA cells (correlation analysis, outlier detection)
3. Execute preprocessing pipeline
4. Train models (LR, RF, XGBoost)
5. Evaluate threshold strategies
6. Save final model artifacts

Driver Dataset Analysis

# Open Car_Claims_Analysis.ipynb
1. Load data (Car_Insurance_Claim.csv)
2. Run EDA cells (risk distributions, missingness analysis)
3. Execute preprocessing pipeline
4. Train models (LR, RF)
5. Compare confusion matrices
6. Review final recommendations

🛠️ Technologies Used

Category Tools
Languages Python 3.8+
Data Processing Pandas, NumPy
Visualization Matplotlib, Seaborn
Machine Learning scikit-learn (LogisticRegression, RandomForestClassifier, Pipeline, ColumnTransformer)
Gradient Boosting XGBoost
Statistical Analysis SciPy (chi2_contingency, Cramér’s V)
Model Persistence joblib
Development Jupyter Notebook
Version Control Git, GitHub

👥 Contributors

Bruno Ndiba Mbwaye Roy

AI/ML Portfolio Project

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

**⭐ Star this repo if you found it helpful!**